charles/monoloco
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Lint (#50)

Browse Source 
- Add continuous integration
- Add Versioneer
- Refactor of preprocessing
- Add tables of evaluation
This commit is contained in:
Lorenzo Bertoni 2021-04-22 15:43:51 +02:00 committed by GitHub
parent 6299859d95
commit 934622bc81
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
61 changed files with 3507 additions and 2112 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
.gitattributes vendored Normal file
View File

@ -0,0 +1 @@
monoloco/_version.py export-subst

81
.github/workflows/tests.yml vendored Normal file
View File

@ -0,0 +1,81 @@
# Adapted from https://github.com/openpifpaf/openpifpaf/blob/main/.github/workflows/test.yml,
#which is: 'Copyright 2019-2021 by Sven Kreiss and contributors. All rights reserved.'
# and licensed under GNU AGPLv3
name: Tests
on: [push, pull_request]
jobs:
build:
runs-on: ${{ matrix.os }}
strategy:
matrix:
include:
- os: ubuntu-latest
python: 3.7
torch: 1.7.1+cpu
torchvision: 0.8.2+cpu
torch-source: https://download.pytorch.org/whl/torch_stable.html
- os: ubuntu-latest
python: 3.8
torch: 1.7.1+cpu
torchvision: 0.8.2+cpu
torch-source: https://download.pytorch.org/whl/cpu/torch_stable.html
- os: macos-latest
python: 3.7
torch: 1.7.1
torchvision: 0.8.2
torch-source: https://download.pytorch.org/whl/torch_stable.html
- os: macos-latest
python: 3.8
torch: 1.7.1
torchvision: 0.8.2
torch-source: https://download.pytorch.org/whl/torch_stable.html
- os: windows-latest
python: 3.7
torch: 1.7.1+cpu
torchvision: 0.8.2+cpu
torch-source: https://download.pytorch.org/whl/torch_stable.html
steps:
- uses: actions/checkout@v2
- name: Set up Python ${{ matrix.python }}
if: ${{ !matrix.conda }}
uses: actions/setup-python@v2
with:
python-version: ${{ matrix.python }}
- name: Set up Conda
if: matrix.conda
uses: s-weigand/setup-conda@v1
with:
update-conda: true
python-version: ${{ matrix.python }}
conda-channels: anaconda, conda-forge
- run: conda --version
if: matrix.conda
- run: which python
if: matrix.conda
- run: python --version
- name: Install
run: |
python -m pip install --upgrade pip setuptools
python -m pip install -e ".[test]"
- name: Print environment
run: |
python -m pip freeze
python --version
python -c "import monoloco; print(monoloco.__version__)"
- name: Lint monoloco
run: |
pylint monoloco --disable=fixme
- name: Lint tests
if: matrix.os != 'windows-latest' # because of path separator
run: |
pylint tests/*.py --disable=fixme
- name: Test
run: |
pytest -vv

2
.gitignore vendored
View File

@ -1,5 +1,5 @@
.idea/ .idea/
data/ data
.DS_store .DS_store
__pycache__ __pycache__
monoloco/*.pyc monoloco/*.pyc

26
.pylintrc
View File

@ -1,26 +0,0 @@
[BASIC]
variable-rgx=[a-z0-9_]{1,30}$ # to accept 2 (dfferent) letters variables
Good-names=xx,dd,zz,hh,ww,pp,kk,lr,w1,w2,w3,mm,im,uv,ax,COV_MIN,CONF_MIN
[TYPECHECK]
disable=import-error,invalid-name,unused-variable,E1102,missing-docstring,useless-object-inheritance,duplicate-code,too-many-arguments,too-many-instance-attributes,too-many-locals,too-few-public-methods,arguments-differ,logging-format-interpolation,import-outside-toplevel
# List of members which are set dynamically and missed by pylint inference
# system, and so shouldn't trigger E1101 when accessed. Python regular
# expressions are accepted.
generated-members=numpy.*,torch.*,cv2.*
ignored-modules=nuscenes, tabulate, cv2
[FORMAT]
max-line-length=120

11
.travis.yml
View File

@ -1,11 +0,0 @@
dist: xenial
language: python
python:
- "3.6"
- "3.7"
install:
- pip install --upgrade pip setuptools
- pip install ".[test]"
script:
- pylint monoloco --disable=unused-variable,fixme
- pytest -v

13
LICENSE
View File

@ -1,4 +1,4 @@
Copyright 2018-2021 by EPFL/VITA. All rights reserved. Copyright 2018-2021 by Lorenzo Bertoni and contributors. All rights reserved.
This project and all its files are licensed under This project and all its files are licensed under
GNU AGPLv3 or later version. GNU AGPLv3 or later version.
@ -7,3 +7,14 @@ If this license is not suitable for your business or project
please contact EPFL-TTO (https://tto.epfl.ch/) for a full commercial license. please contact EPFL-TTO (https://tto.epfl.ch/) for a full commercial license.
This software may not be used to harm any person deliberately or for any military application. This software may not be used to harm any person deliberately or for any military application.
The following files are based on the OpenPifPaf project which is
"Copyright 2019-2021 by Sven Kreiss and contributors. All rights reserved." and licensed under GNU AGPLv3.
- tests/test_train_mono.py
- tests/test_train_stereo.py
- monoloco/visuals/pifpaf_show.py
- monoloco/train/losses.py
- monoloco/predict.py
-.github/workflows/tests.yml

661
LICENSE.AGPL
View File

@ -1,661 +0,0 @@
GNU AFFERO GENERAL PUBLIC LICENSE
Version 3, 19 November 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU Affero General Public License is a free, copyleft license for
software and other kinds of works, specifically designed to ensure
cooperation with the community in the case of network server software.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
our General Public Licenses are intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
Developers that use our General Public Licenses protect your rights
with two steps: (1) assert copyright on the software, and (2) offer
you this License which gives you legal permission to copy, distribute
and/or modify the software.
A secondary benefit of defending all users' freedom is that
improvements made in alternate versions of the program, if they
receive widespread use, become available for other developers to
incorporate. Many developers of free software are heartened and
encouraged by the resulting cooperation. However, in the case of
software used on network servers, this result may fail to come about.
The GNU General Public License permits making a modified version and
letting the public access it on a server without ever releasing its
source code to the public.
The GNU Affero General Public License is designed specifically to
ensure that, in such cases, the modified source code becomes available
to the community. It requires the operator of a network server to
provide the source code of the modified version running there to the
users of that server. Therefore, public use of a modified version, on
a publicly accessible server, gives the public access to the source
code of the modified version.
An older license, called the Affero General Public License and
published by Affero, was designed to accomplish similar goals. This is
a different license, not a version of the Affero GPL, but Affero has
released a new version of the Affero GPL which permits relicensing under
this license.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU Affero General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Remote Network Interaction; Use with the GNU General Public License.
Notwithstanding any other provision of this License, if you modify the
Program, your modified version must prominently offer all users
interacting with it remotely through a computer network (if your version
supports such interaction) an opportunity to receive the Corresponding
Source of your version by providing access to the Corresponding Source
from a network server at no charge, through some standard or customary
means of facilitating copying of software. This Corresponding Source
shall include the Corresponding Source for any work covered by version 3
of the GNU General Public License that is incorporated pursuant to the
following paragraph.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the work with which it is combined will remain governed by version
3 of the GNU General Public License.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU Affero General Public License from time to time. Such new versions
will be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU Affero General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU Affero General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU Affero General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If your software can interact with users remotely through a computer
network, you should also make sure that it provides a way for users to
get its source. For example, if your program is a web application, its
interface could display a "Source" link that leads users to an archive
of the code. There are many ways you could offer source, and different
solutions will be better for different programs; see section 13 for the
specific requirements.
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU AGPL, see
<http://www.gnu.org/licenses/>.

2
MANIFEST.in Normal file
View File

@ -0,0 +1,2 @@
include versioneer.py
include monoloco/_version.py

64
README.md
View File

@ -1,9 +1,12 @@
# Monoloco library &nbsp;&nbsp; &nbsp; [![Downloads](https://pepy.tech/badge/monoloco)](https://pepy.tech/project/monoloco) # Monoloco library &nbsp;&nbsp; [![Downloads](https://pepy.tech/badge/monoloco)](https://pepy.tech/project/monoloco)
Continuously tested on Linux, MacOS and Windows: [![Tests](https://github.com/vita-epfl/monoloco/workflows/Tests/badge.svg?branch=main)](https://github.com/vita-epfl/monoloco/actions?query=workflow%3ATests)
<img src="docs/monoloco.gif" alt="gif" /> <img src="docs/monoloco.gif" alt="gif" />
This library is based on three research projects for monocular/stereo 3D human localization (detection), body orientation, and social distancing. Check the __video teaser__ of the library on [__YouTube__](https://www.youtube.com/watch?v=O5zhzi8mwJ4). This library is based on three research projects for monocular/stereo 3D human localization (detection), body orientation, and social distancing. Check the __video teaser__ of the library on [__YouTube__](https://www.youtube.com/watch?v=O5zhzi8mwJ4).
--- ---
@ -114,9 +117,9 @@ If you provide a ground-truth json file to compare the predictions of the networ
For an example image, run the following command: For an example image, run the following command:
``` ```sh
python -m monoloco.run predict docs/002282.png \ python -m monoloco.run predict docs/002282.png \
--path_gt <to match results with ground-truths> \ --path_gt names-kitti-200615-1022.json \
-o <output directory> \ -o <output directory> \
--long-edge <rescale the image by providing dimension of long side> --long-edge <rescale the image by providing dimension of long side>
--n_dropout <50 to include epistemic uncertainty, 0 otherwise> --n_dropout <50 to include epistemic uncertainty, 0 otherwise>
@ -129,7 +132,7 @@ To show all the instances estimated by MonoLoco add the argument `show_all` to t
![predict_all](docs/out_002282.png.multi_all.jpg) ![predict_all](docs/out_002282.png.multi_all.jpg)
It is also possible to run [openpifpaf](https://github.com/vita-epfl/openpifpaf) directly It is also possible to run [openpifpaf](https://github.com/vita-epfl/openpifpaf) directly
by usingt `--mode keypoints`. All the other pifpaf arguments are also supported by using `--mode keypoints`. All the other pifpaf arguments are also supported
and can be checked with `python -m monoloco.run predict --help`. and can be checked with `python -m monoloco.run predict --help`.
![predict](docs/out_002282_pifpaf.jpg) ![predict](docs/out_002282_pifpaf.jpg)
@ -199,7 +202,7 @@ python -m monoloco.run train --joints data/arrays/joints-kitti-201202-1743.json
While for the MonStereo ones just change the input joints and add `--mode stereo` While for the MonStereo ones just change the input joints and add `--mode stereo`
``` ```
python3 -m monoloco.run train --joints data/arrays/joints-kitti-201202-1022.json --mode stereo python3 -m monoloco.run train --lr 0.002 --joints data/arrays/joints-kitti-201202-1022.json --mode stereo
``` ```
If you are interested in the original results of the MonoLoco ICCV article (now improved with MonoLoco++), please refer to the tag v0.4.9 in this repository. If you are interested in the original results of the MonoLoco ICCV article (now improved with MonoLoco++), please refer to the tag v0.4.9 in this repository.
@ -219,40 +222,58 @@ The code supports this option (by running the predict script and using `--mode p
### Data structure ### Data structure
data data
├── arrays ├── outputs
├── models ├── arrays
├── kitti ├── kitti
├── logs
├── output
Run the following inside monoloco repository: Run the following inside monoloco repository:
``` ```
mkdir data mkdir data
cd data cd data
mkdir arrays models kitti logs output mkdir outputs arrays kitti
``` ```
### Kitti Dataset ### Kitti Dataset
Annotations from a pose detector needs to be stored in a folder. With PifPaf: Download kitti images (from left and right cameras), ground-truth files (labels), and calibration files from their [website](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d) and save them inside the `data` folder as shown below.
``` data
├── kitti
├── gt
├── calib
├── images
├── images_right
The network takes as inputs 2D keypoints annotations. To create them run PifPaf over the saved images:
```sh
python -m openpifpaf.predict \ python -m openpifpaf.predict \
--glob "<kitti images directory>/*.png" \ --glob "data/kitti/images/*.png" \
--json-output <directory to contain predictions> \ --json-output <directory to contain predictions> \
--checkpoint=shufflenetv2k30 \ --checkpoint=shufflenetv2k30 \
--instance-threshold=0.05 --seed-threshold 0.05 --force-complete-pose --instance-threshold=0.05 --seed-threshold 0.05 --force-complete-pose
``` ```
Once the step is complete, the below commands transform all the annotations into a single json file that will used for training. **Horizontal flipping**
To augment the dataset, we apply horizontal flipping on the detected poses. To include small variations in the pose, we use the poses from the right-camera (the dataset uses a stereo camera). As there are no labels for the right camera, the code automatically correct the ground truth depth by taking into account the camera baseline.
To obtain these poses, run pifpaf also on the folder of right images. Make sure to save annotations into a different folder, and call the right folder: `<NameOfTheLeftFolder>_right`
**Recall**
To maximize the recall (at the cost of the computational time), it's possible to upscale the images with the command `--long_edge 2500` (\~scale 2).
Once this step is complete, the below commands transform all the annotations into a single json file that will used for training.
For MonoLoco++: For MonoLoco++:
``` ```sh
python -m monoloco.run prep --dir_ann <directory that contains annotations> python -m monoloco.run prep --dir_ann <directory that contains annotations>
``` ```
For MonStereo: For MonStereo:
``` ```sh
python -m monoloco.run prep --mode stereo --dir_ann <directory that contains annotations> python -m monoloco.run prep --mode stereo --dir_ann <directory that contains left annotations>
``` ```
### Collective Activity Dataset ### Collective Activity Dataset
@ -324,18 +345,19 @@ and save them into `data/kitti/monodepth`
To include also geometric baselines and MonoLoco, download a monoloco model, save it in `data/models`, and add the flag ``--baselines`` to the evaluation command To include also geometric baselines and MonoLoco, download a monoloco model, save it in `data/models`, and add the flag ``--baselines`` to the evaluation command
The evaluation file will run the model over all the annotations and compare the results with KITTI ground-truth and the downloaded baselines. For this run: The evaluation file will run the model over all the annotations and compare the results with KITTI ground-truth and the downloaded baselines. For this run:
``` ```
python -m monoloco.run eval \ python -m monoloco.run eval \
--dir_ann <annotation directory> \ --dir_ann <annotation directory> \
--model <model path> \ --model #TODO \
--generate \ --generate \
--save \ --save \
```` ````
For stereo results add `--mode stereo` and select `--model #TODO. Below, the resulting table of results and an example of the saved figures.
<img src="docs/results.jpg" width="700"/>
<img src="docs/results_stereo.jpg" width="550"/> <img src="docs/results_monstereo.jpg" width="700"/>
### Relative Average Precision Localization: RALP-5% (MonStereo) ### Relative Average Precision Localization: RALP-5% (MonStereo)

BIN
docs/quantitative_mono.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 295 KiB

BIN
docs/results.jpg Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 806 KiB

BIN
docs/results_monoloco_pp.jpg
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 385 KiB

BIN
docs/results_monstereo.jpg Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 348 KiB

BIN
docs/results_stereo.jpg
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 368 KiB

4
monoloco/__init__.py
View File

@ -3,4 +3,6 @@
Open implementation of MonoLoco / MonoLoco++ / MonStereo Open implementation of MonoLoco / MonoLoco++ / MonStereo
""" """
__version__ = '0.5.0' from ._version import get_versions
__version__ = get_versions()['version']
del get_versions

527
monoloco/_version.py Normal file
View File

@ -0,0 +1,527 @@
# This file helps to compute a version number in source trees obtained from
# git-archive tarball (such as those provided by githubs download-from-tag
# feature). Distribution tarballs (built by setup.py sdist) and build
# directories (produced by setup.py build) will contain a much shorter file
# that just contains the computed version number.
# This file is released into the public domain. Generated by
# versioneer-0.19 (https://github.com/python-versioneer/python-versioneer)
# pylint: skip-file
"""Git implementation of _version.py."""
import errno
import os
import re
import subprocess
import sys
def get_keywords():
"""Get the keywords needed to look up the version information."""
# these strings will be replaced by git during git-archive.
# setup.py/versioneer.py will grep for the variable names, so they must
# each be defined on a line of their own. _version.py will just call
# get_keywords().
git_refnames = "$Format:%d$"
git_full = "$Format:%H$"
git_date = "$Format:%ci$"
keywords = {"refnames": git_refnames, "full": git_full, "date": git_date}
return keywords
class VersioneerConfig:
"""Container for Versioneer configuration parameters."""
def get_config():
"""Create, populate and return the VersioneerConfig() object."""
# these strings are filled in when 'setup.py versioneer' creates
# _version.py
cfg = VersioneerConfig()
cfg.VCS = "git"
cfg.style = "pep440"
cfg.tag_prefix = "v"
cfg.parentdir_prefix = "None"
cfg.versionfile_source = "monoloco/_version.py"
cfg.verbose = False
return cfg
class NotThisMethod(Exception):
"""Exception raised if a method is not valid for the current scenario."""
LONG_VERSION_PY = {}
HANDLERS = {}
def register_vcs_handler(vcs, method): # decorator
"""Create decorator to mark a method as the handler of a VCS."""
def decorate(f):
"""Store f in HANDLERS[vcs][method]."""
if vcs not in HANDLERS:
HANDLERS[vcs] = {}
HANDLERS[vcs][method] = f
return f
return decorate
def run_command(commands, args, cwd=None, verbose=False, hide_stderr=False,
env=None):
"""Call the given command(s)."""
assert isinstance(commands, list)
p = None
for c in commands:
try:
dispcmd = str([c] + args)
# remember shell=False, so use git.cmd on windows, not just git
p = subprocess.Popen([c] + args, cwd=cwd, env=env,
stdout=subprocess.PIPE,
stderr=(subprocess.PIPE if hide_stderr
else None))
break
except EnvironmentError:
e = sys.exc_info()[1]
if e.errno == errno.ENOENT:
continue
if verbose:
print("unable to run %s" % dispcmd)
print(e)
return None, None
else:
if verbose:
print("unable to find command, tried %s" % (commands,))
return None, None
stdout = p.communicate()[0].strip().decode()
if p.returncode != 0:
if verbose:
print("unable to run %s (error)" % dispcmd)
print("stdout was %s" % stdout)
return None, p.returncode
return stdout, p.returncode
def versions_from_parentdir(parentdir_prefix, root, verbose):
"""Try to determine the version from the parent directory name.
Source tarballs conventionally unpack into a directory that includes both
the project name and a version string. We will also support searching up
two directory levels for an appropriately named parent directory
"""
rootdirs = []
for i in range(3):
dirname = os.path.basename(root)
if dirname.startswith(parentdir_prefix):
return {"version": dirname[len(parentdir_prefix):],
"full-revisionid": None,
"dirty": False, "error": None, "date": None}
else:
rootdirs.append(root)
root = os.path.dirname(root) # up a level
if verbose:
print("Tried directories %s but none started with prefix %s" %
(str(rootdirs), parentdir_prefix))
raise NotThisMethod("rootdir doesn't start with parentdir_prefix")
@register_vcs_handler("git", "get_keywords")
def git_get_keywords(versionfile_abs):
"""Extract version information from the given file."""
# the code embedded in _version.py can just fetch the value of these
# keywords. When used from setup.py, we don't want to import _version.py,
# so we do it with a regexp instead. This function is not used from
# _version.py.
keywords = {}
try:
f = open(versionfile_abs, "r")
for line in f.readlines():
if line.strip().startswith("git_refnames ="):
mo = re.search(r'=\s*"(.*)"', line)
if mo:
keywords["refnames"] = mo.group(1)
if line.strip().startswith("git_full ="):
mo = re.search(r'=\s*"(.*)"', line)
if mo:
keywords["full"] = mo.group(1)
if line.strip().startswith("git_date ="):
mo = re.search(r'=\s*"(.*)"', line)
if mo:
keywords["date"] = mo.group(1)
f.close()
except EnvironmentError:
pass
return keywords
@register_vcs_handler("git", "keywords")
def git_versions_from_keywords(keywords, tag_prefix, verbose):
"""Get version information from git keywords."""
if not keywords:
raise NotThisMethod("no keywords at all, weird")
date = keywords.get("date")
if date is not None:
# Use only the last line. Previous lines may contain GPG signature
# information.
date = date.splitlines()[-1]
# git-2.2.0 added "%cI", which expands to an ISO-8601 -compliant
# datestamp. However we prefer "%ci" (which expands to an "ISO-8601
# -like" string, which we must then edit to make compliant), because
# it's been around since git-1.5.3, and it's too difficult to
# discover which version we're using, or to work around using an
# older one.
date = date.strip().replace(" ", "T", 1).replace(" ", "", 1)
refnames = keywords["refnames"].strip()
if refnames.startswith("$Format"):
if verbose:
print("keywords are unexpanded, not using")
raise NotThisMethod("unexpanded keywords, not a git-archive tarball")
refs = set([r.strip() for r in refnames.strip("()").split(",")])
# starting in git-1.8.3, tags are listed as "tag: foo-1.0" instead of
# just "foo-1.0". If we see a "tag: " prefix, prefer those.
TAG = "tag: "
tags = set([r[len(TAG):] for r in refs if r.startswith(TAG)])
if not tags:
# Either we're using git < 1.8.3, or there really are no tags. We use
# a heuristic: assume all version tags have a digit. The old git %d
# expansion behaves like git log --decorate=short and strips out the
# refs/heads/ and refs/tags/ prefixes that would let us distinguish
# between branches and tags. By ignoring refnames without digits, we
# filter out many common branch names like "release" and
# "stabilization", as well as "HEAD" and "master".
tags = set([r for r in refs if re.search(r'\d', r)])
if verbose:
print("discarding '%s', no digits" % ",".join(refs - tags))
if verbose:
print("likely tags: %s" % ",".join(sorted(tags)))
for ref in sorted(tags):
# sorting will prefer e.g. "2.0" over "2.0rc1"
if ref.startswith(tag_prefix):
r = ref[len(tag_prefix):]
if verbose:
print("picking %s" % r)
return {"version": r,
"full-revisionid": keywords["full"].strip(),
"dirty": False, "error": None,
"date": date}
# no suitable tags, so version is "0+unknown", but full hex is still there
if verbose:
print("no suitable tags, using unknown + full revision id")
return {"version": "0+unknown",
"full-revisionid": keywords["full"].strip(),
"dirty": False, "error": "no suitable tags", "date": None}
@register_vcs_handler("git", "pieces_from_vcs")
def git_pieces_from_vcs(tag_prefix, root, verbose, run_command=run_command):
"""Get version from 'git describe' in the root of the source tree.
This only gets called if the git-archive 'subst' keywords were *not*
expanded, and _version.py hasn't already been rewritten with a short
version string, meaning we're inside a checked out source tree.
"""
GITS = ["git"]
if sys.platform == "win32":
GITS = ["git.cmd", "git.exe"]
out, rc = run_command(GITS, ["rev-parse", "--git-dir"], cwd=root,
hide_stderr=True)
if rc != 0:
if verbose:
print("Directory %s not under git control" % root)
raise NotThisMethod("'git rev-parse --git-dir' returned error")
# if there is a tag matching tag_prefix, this yields TAG-NUM-gHEX[-dirty]
# if there isn't one, this yields HEX[-dirty] (no NUM)
describe_out, rc = run_command(GITS, ["describe", "--tags", "--dirty",
"--always", "--long",
"--match", "%s*" % tag_prefix],
cwd=root)
# --long was added in git-1.5.5
if describe_out is None:
raise NotThisMethod("'git describe' failed")
describe_out = describe_out.strip()
full_out, rc = run_command(GITS, ["rev-parse", "HEAD"], cwd=root)
if full_out is None:
raise NotThisMethod("'git rev-parse' failed")
full_out = full_out.strip()
pieces = {}
pieces["long"] = full_out
pieces["short"] = full_out[:7] # maybe improved later
pieces["error"] = None
# parse describe_out. It will be like TAG-NUM-gHEX[-dirty] or HEX[-dirty]
# TAG might have hyphens.
git_describe = describe_out
# look for -dirty suffix
dirty = git_describe.endswith("-dirty")
pieces["dirty"] = dirty
if dirty:
git_describe = git_describe[:git_describe.rindex("-dirty")]
# now we have TAG-NUM-gHEX or HEX
if "-" in git_describe:
# TAG-NUM-gHEX
mo = re.search(r'^(.+)-(\d+)-g([0-9a-f]+)$', git_describe)
if not mo:
# unparseable. Maybe git-describe is misbehaving?
pieces["error"] = ("unable to parse git-describe output: '%s'"
% describe_out)
return pieces
# tag
full_tag = mo.group(1)
if not full_tag.startswith(tag_prefix):
if verbose:
fmt = "tag '%s' doesn't start with prefix '%s'"
print(fmt % (full_tag, tag_prefix))
pieces["error"] = ("tag '%s' doesn't start with prefix '%s'"
% (full_tag, tag_prefix))
return pieces
pieces["closest-tag"] = full_tag[len(tag_prefix):]
# distance: number of commits since tag
pieces["distance"] = int(mo.group(2))
# commit: short hex revision ID
pieces["short"] = mo.group(3)
else:
# HEX: no tags
pieces["closest-tag"] = None
count_out, rc = run_command(GITS, ["rev-list", "HEAD", "--count"],
cwd=root)
pieces["distance"] = int(count_out) # total number of commits
# commit date: see ISO-8601 comment in git_versions_from_keywords()
date = run_command(GITS, ["show", "-s", "--format=%ci", "HEAD"],
cwd=root)[0].strip()
# Use only the last line. Previous lines may contain GPG signature
# information.
date = date.splitlines()[-1]
pieces["date"] = date.strip().replace(" ", "T", 1).replace(" ", "", 1)
return pieces
def plus_or_dot(pieces):
"""Return a + if we don't already have one, else return a ."""
if "+" in pieces.get("closest-tag", ""):
return "."
return "+"
def render_pep440(pieces):
"""Build up version string, with post-release "local version identifier".
Our goal: TAG[+DISTANCE.gHEX[.dirty]] . Note that if you
get a tagged build and then dirty it, you'll get TAG+0.gHEX.dirty
Exceptions:
1: no tags. git_describe was just HEX. 0+untagged.DISTANCE.gHEX[.dirty]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
rendered += plus_or_dot(pieces)
rendered += "%d.g%s" % (pieces["distance"], pieces["short"])
if pieces["dirty"]:
rendered += ".dirty"
else:
# exception #1
rendered = "0+untagged.%d.g%s" % (pieces["distance"],
pieces["short"])
if pieces["dirty"]:
rendered += ".dirty"
return rendered
def render_pep440_pre(pieces):
"""TAG[.post0.devDISTANCE] -- No -dirty.
Exceptions:
1: no tags. 0.post0.devDISTANCE
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"]:
rendered += ".post0.dev%d" % pieces["distance"]
else:
# exception #1
rendered = "0.post0.dev%d" % pieces["distance"]
return rendered
def render_pep440_post(pieces):
"""TAG[.postDISTANCE[.dev0]+gHEX] .
The ".dev0" means dirty. Note that .dev0 sorts backwards
(a dirty tree will appear "older" than the corresponding clean one),
but you shouldn't be releasing software with -dirty anyways.
Exceptions:
1: no tags. 0.postDISTANCE[.dev0]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
rendered += ".post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
rendered += plus_or_dot(pieces)
rendered += "g%s" % pieces["short"]
else:
# exception #1
rendered = "0.post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
rendered += "+g%s" % pieces["short"]
return rendered
def render_pep440_old(pieces):
"""TAG[.postDISTANCE[.dev0]] .
The ".dev0" means dirty.
Exceptions:
1: no tags. 0.postDISTANCE[.dev0]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
rendered += ".post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
else:
# exception #1
rendered = "0.post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
return rendered
def render_git_describe(pieces):
"""TAG[-DISTANCE-gHEX][-dirty].
Like 'git describe --tags --dirty --always'.
Exceptions:
1: no tags. HEX[-dirty] (note: no 'g' prefix)
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"]:
rendered += "-%d-g%s" % (pieces["distance"], pieces["short"])
else:
# exception #1
rendered = pieces["short"]
if pieces["dirty"]:
rendered += "-dirty"
return rendered
def render_git_describe_long(pieces):
"""TAG-DISTANCE-gHEX[-dirty].
Like 'git describe --tags --dirty --always -long'.
The distance/hash is unconditional.
Exceptions:
1: no tags. HEX[-dirty] (note: no 'g' prefix)
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
rendered += "-%d-g%s" % (pieces["distance"], pieces["short"])
else:
# exception #1
rendered = pieces["short"]
if pieces["dirty"]:
rendered += "-dirty"
return rendered
def render(pieces, style):
"""Render the given version pieces into the requested style."""
if pieces["error"]:
return {"version": "unknown",
"full-revisionid": pieces.get("long"),
"dirty": None,
"error": pieces["error"],
"date": None}
if not style or style == "default":
style = "pep440" # the default
if style == "pep440":
rendered = render_pep440(pieces)
elif style == "pep440-pre":
rendered = render_pep440_pre(pieces)
elif style == "pep440-post":
rendered = render_pep440_post(pieces)
elif style == "pep440-old":
rendered = render_pep440_old(pieces)
elif style == "git-describe":
rendered = render_git_describe(pieces)
elif style == "git-describe-long":
rendered = render_git_describe_long(pieces)
else:
raise ValueError("unknown style '%s'" % style)
return {"version": rendered, "full-revisionid": pieces["long"],
"dirty": pieces["dirty"], "error": None,
"date": pieces.get("date")}
def get_versions():
"""Get version information or return default if unable to do so."""
# I am in _version.py, which lives at ROOT/VERSIONFILE_SOURCE. If we have
# __file__, we can work backwards from there to the root. Some
# py2exe/bbfreeze/non-CPython implementations don't do __file__, in which
# case we can only use expanded keywords.
cfg = get_config()
verbose = cfg.verbose
try:
return git_versions_from_keywords(get_keywords(), cfg.tag_prefix,
verbose)
except NotThisMethod:
pass
try:
root = os.path.realpath(__file__)
# versionfile_source is the relative path from the top of the source
# tree (where the .git directory might live) to this file. Invert
# this to find the root from __file__.
for i in cfg.versionfile_source.split('/'):
root = os.path.dirname(root)
except NameError:
return {"version": "0+unknown", "full-revisionid": None,
"dirty": None,
"error": "unable to find root of source tree",
"date": None}
try:
pieces = git_pieces_from_vcs(cfg.tag_prefix, root, verbose)
return render(pieces, cfg.style)
except NotThisMethod:
pass
try:
if cfg.parentdir_prefix:
return versions_from_parentdir(cfg.parentdir_prefix, root, verbose)
except NotThisMethod:
pass
return {"version": "0+unknown", "full-revisionid": None,
"dirty": None,
"error": "unable to compute version", "date": None}

2
monoloco/activity.py
View File

@ -123,7 +123,7 @@ def show_social(args, image_t, output_path, annotations, dic_out):
# Draw keypoints and orientation # Draw keypoints and orientation
if 'front' in args.output_types: if 'front' in args.output_types:
keypoint_sets, scores = get_pifpaf_outputs(annotations) keypoint_sets, _ = get_pifpaf_outputs(annotations)
uv_centers = dic_out['uv_heads'] uv_centers = dic_out['uv_heads']
sizes = [abs(dic_out['uv_heads'][idx][1] - uv_s[1]) / 1.5 for idx, uv_s in sizes = [abs(dic_out['uv_heads'][idx][1] - uv_s[1]) / 1.5 for idx, uv_s in
enumerate(dic_out['uv_shoulders'])] enumerate(dic_out['uv_shoulders'])]

28
monoloco/eval/eval_activity.py
View File

@ -2,17 +2,23 @@
import os import os
import glob import glob
import csv import csv
import copy
from collections import defaultdict from collections import defaultdict
import numpy as np import numpy as np
import torch import torch
from PIL import Image from PIL import Image
from sklearn.metrics import accuracy_score try:
from sklearn.metrics import accuracy_score
ACCURACY_SCORE = copy.copy(accuracy_score)
except ImportError:
ACCURACY_SCORE = None
from monoloco.network import Loco from ..prep import factory_file
from monoloco.network.process import factory_for_gt, preprocess_pifpaf from ..network import Loco
from monoloco.activity import social_interactions from ..network.process import factory_for_gt, preprocess_pifpaf
from monoloco.utils import open_annotations, get_iou_matches, get_difficulty from ..activity import social_interactions
from ..utils import open_annotations, get_iou_matches, get_difficulty
class ActivityEvaluator: class ActivityEvaluator:
@ -79,7 +85,7 @@ class ActivityEvaluator:
im_size = image.size im_size = image.size
assert len(im_size) > 1, "image with frame0001 not available" assert len(im_size) > 1, "image with frame0001 not available"
for idx, im_path in enumerate(images): for im_path in images:
# Collect PifPaf files and calibration # Collect PifPaf files and calibration
basename = os.path.basename(im_path) basename = os.path.basename(im_path)
@ -101,14 +107,12 @@ class ActivityEvaluator:
self.estimate_activity(dic_out, matches, ys_gt, categories=categories) self.estimate_activity(dic_out, matches, ys_gt, categories=categories)
# Print Results # Print Results
acc = accuracy_score(self.all_gt[seq], self.all_pred[seq]) acc = ACCURACY_SCORE(self.all_gt[seq], self.all_pred[seq])
print(f"Accuracy of category {seq}: {100*acc:.2f}%") print(f"Accuracy of category {seq}: {100*acc:.2f}%")
cout_results(self.cnt, self.all_gt, self.all_pred, categories=self.sequences) cout_results(self.cnt, self.all_gt, self.all_pred, categories=self.sequences)
def eval_kitti(self): def eval_kitti(self):
"""Parse KITTI Dataset and predict if people are talking or not""" """Parse KITTI Dataset and predict if people are talking or not"""
from ..utils import factory_file
files = glob.glob(self.dir_data + '/*.txt') files = glob.glob(self.dir_data + '/*.txt')
# files = [self.dir_gt_kitti + '/001782.txt'] # files = [self.dir_gt_kitti + '/001782.txt']
assert files, "Empty directory" assert files, "Empty directory"
@ -118,7 +122,7 @@ class ActivityEvaluator:
# Collect PifPaf files and calibration # Collect PifPaf files and calibration
basename, _ = os.path.splitext(os.path.basename(file)) basename, _ = os.path.splitext(os.path.basename(file))
path_calib = os.path.join(self.dir_kk, basename + '.txt') path_calib = os.path.join(self.dir_kk, basename + '.txt')
annotations, kk, tt = factory_file(path_calib, self.dir_ann, basename) annotations, kk, _ = factory_file(path_calib, self.dir_ann, basename)
# Collect corresponding gt files (ys_gt: 1 or 0) # Collect corresponding gt files (ys_gt: 1 or 0)
path_gt = os.path.join(self.dir_data, basename + '.txt') path_gt = os.path.join(self.dir_data, basename + '.txt')
@ -149,7 +153,7 @@ class ActivityEvaluator:
self.cnt['gt'][key] += 1 self.cnt['gt'][key] += 1
self.cnt['gt']['all'] += 1 self.cnt['gt']['all'] += 1
for i_m, (idx, idx_gt) in enumerate(matches): for (idx, idx_gt) in matches:
# Select keys to update results for Collective or KITTI # Select keys to update results for Collective or KITTI
keys = ('all', categories[idx_gt]) keys = ('all', categories[idx_gt])
@ -184,7 +188,7 @@ def parse_gt_collective(dir_data, seq, path_pif):
with open(path, "r") as ff: with open(path, "r") as ff:
reader = csv.reader(ff, delimiter='\t') reader = csv.reader(ff, delimiter='\t')
dic_frames = defaultdict(lambda: defaultdict(list)) dic_frames = defaultdict(lambda: defaultdict(list))
for idx, line in enumerate(reader): for line in reader:
box = convert_box(line[1:5]) box = convert_box(line[1:5])
cat = convert_category(line[5]) cat = convert_category(line[5])
dic_frames[line[0]]['boxes'].append(box) dic_frames[line[0]]['boxes'].append(box)

36
monoloco/eval/eval_kitti.py
View File

@ -7,13 +7,20 @@ Evaluate MonStereo code on KITTI dataset using ALE metric
import os import os
import math import math
import logging import logging
import copy
import datetime import datetime
from collections import defaultdict from collections import defaultdict
from tabulate import tabulate import numpy as np
try:
import tabulate
TABULATE = copy.copy(tabulate.tabulate)
except ImportError:
TABULATE = None
from ..utils import get_iou_matches, get_task_error, get_pixel_error, check_conditions, \ from ..utils import get_iou_matches, get_task_error, get_pixel_error, check_conditions, \
get_difficulty, split_training, parse_ground_truth, get_iou_matches_matrix, average, find_cluster get_difficulty, split_training, get_iou_matches_matrix, average, find_cluster
from ..prep import parse_ground_truth
from ..visuals import show_results, show_spread, show_task_error, show_box_plot from ..visuals import show_results, show_spread, show_task_error, show_box_plot
@ -29,7 +36,7 @@ class EvalKitti:
METHODS_STEREO = ['3dop', 'psf', 'pseudo-lidar', 'e2e', 'oc-stereo'] METHODS_STEREO = ['3dop', 'psf', 'pseudo-lidar', 'e2e', 'oc-stereo']
BASELINES = ['task_error', 'pixel_error'] BASELINES = ['task_error', 'pixel_error']
HEADERS = ('method', '<0.5', '<1m', '<2m', 'easy', 'moderate', 'hard', 'all') HEADERS = ('method', '<0.5', '<1m', '<2m', 'easy', 'moderate', 'hard', 'all')
CATEGORIES = ('pedestrian',) CATEGORIES = ('pedestrian',) # extendable with person_sitting and/or cyclists
methods = OUR_METHODS + METHODS_MONO + METHODS_STEREO methods = OUR_METHODS + METHODS_MONO + METHODS_STEREO
# Set directories # Set directories
@ -39,8 +46,7 @@ class EvalKitti:
path_val = os.path.join('splits', 'kitti_val.txt') path_val = os.path.join('splits', 'kitti_val.txt')
dir_logs = os.path.join('data', 'logs') dir_logs = os.path.join('data', 'logs')
assert os.path.exists(dir_logs), "No directory to save final statistics" assert os.path.exists(dir_logs), "No directory to save final statistics"
dir_fig = os.path.join('data', 'figures') dir_fig = os.path.join('figures', 'results')
assert os.path.exists(dir_logs), "No directory to save figures"
# Set thresholds to obtain comparable recalls # Set thresholds to obtain comparable recalls
thresh_iou_monoloco = 0.3 thresh_iou_monoloco = 0.3
@ -49,9 +55,10 @@ class EvalKitti:
thresh_conf_base = 0.5 thresh_conf_base = 0.5
def __init__(self, args): def __init__(self, args):
self.mode = args.mode
assert self.mode in ('mono', 'stereo'), "mode not recognized"
self.net = 'monstereo' if self.mode == 'stereo' else 'monoloco_pp'
self.verbose = args.verbose self.verbose = args.verbose
self.net = args.net
self.save = args.save self.save = args.save
self.show = args.show self.show = args.show
@ -110,7 +117,7 @@ class EvalKitti:
methods_out = defaultdict(tuple) # Save all methods for comparison methods_out = defaultdict(tuple) # Save all methods for comparison
# Count ground_truth: # Count ground_truth:
boxes_gt, ys, truncs_gt, occs_gt = out_gt # pylint: disable=unbalanced-tuple-unpacking boxes_gt, _, truncs_gt, occs_gt, _ = out_gt # pylint: disable=unbalanced-tuple-unpacking
for idx, box in enumerate(boxes_gt): for idx, box in enumerate(boxes_gt):
mode = get_difficulty(box, truncs_gt[idx], occs_gt[idx]) mode = get_difficulty(box, truncs_gt[idx], occs_gt[idx])
self.cnt_gt[mode] += 1 self.cnt_gt[mode] += 1
@ -144,10 +151,13 @@ class EvalKitti:
self.show_statistics() self.show_statistics()
def printer(self): def printer(self):
if self.save:
os.makedirs(self.dir_fig, exist_ok=True)
if self.save or self.show: if self.save or self.show:
print('-' * 100)
show_results(self.dic_stats, self.CLUSTERS, self.net, self.dir_fig, show=self.show, save=self.save) show_results(self.dic_stats, self.CLUSTERS, self.net, self.dir_fig, show=self.show, save=self.save)
show_spread(self.dic_stats, self.CLUSTERS, self.net, self.dir_fig, show=self.show, save=self.save) show_spread(self.dic_stats, self.CLUSTERS, self.net, self.dir_fig, show=self.show, save=self.save)
if self.net == 'monstero': if self.net == 'monstereo':
show_box_plot(self.errors, self.CLUSTERS, self.dir_fig, show=self.show, save=self.save) show_box_plot(self.errors, self.CLUSTERS, self.dir_fig, show=self.show, save=self.save)
else: else:
show_task_error(self.dir_fig, show=self.show, save=self.save) show_task_error(self.dir_fig, show=self.show, save=self.save)
@ -201,7 +211,7 @@ class EvalKitti:
def _estimate_error(self, out_gt, out, method): def _estimate_error(self, out_gt, out, method):
"""Estimate localization error""" """Estimate localization error"""
boxes_gt, ys, truncs_gt, occs_gt = out_gt boxes_gt, ys, truncs_gt, occs_gt, _ = out_gt
if method in self.OUR_METHODS: if method in self.OUR_METHODS:
boxes, dds, cat, bis, epis = out boxes, dds, cat, bis, epis = out
@ -363,7 +373,7 @@ class EvalKitti:
for key in all_methods] for key in all_methods]
results = [[key] + alp[idx] + ale[idx] for idx, key in enumerate(all_methods)] results = [[key] + alp[idx] + ale[idx] for idx, key in enumerate(all_methods)]
print(tabulate(results, headers=self.HEADERS)) print(TABULATE(results, headers=self.HEADERS))
print('-' * 90 + '\n') print('-' * 90 + '\n')
def stats_height(self): def stats_height(self):
@ -373,10 +383,8 @@ class EvalKitti:
self.name = name self.name = name
# Iterate over each line of the gt file and save box location and distances # Iterate over each line of the gt file and save box location and distances
out_gt = parse_ground_truth(path_gt, 'pedestrian') out_gt = parse_ground_truth(path_gt, 'pedestrian')
boxes_gt, ys, truncs_gt, occs_gt = out_gt # pylint: disable=unbalanced-tuple-unpacking for label in out_gt[1]:
for label in ys:
heights.append(label[4]) heights.append(label[4])
import numpy as np
tail1, tail2 = np.nanpercentile(np.array(heights), [5, 95]) tail1, tail2 = np.nanpercentile(np.array(heights), [5, 95])
print(average(heights)) print(average(heights))
print(len(heights)) print(len(heights))

27
monoloco/eval/eval_variance.py
View File

@ -1,4 +1,4 @@
# pylint: disable=too-many-statements,too-many-branches # pylint: disable=too-many-statements
"""Joints Analysis: Supplementary material of MonStereo""" """Joints Analysis: Supplementary material of MonStereo"""
@ -11,26 +11,7 @@ import matplotlib.pyplot as plt
from ..utils import find_cluster, average from ..utils import find_cluster, average
from ..visuals.figures import get_distances from ..visuals.figures import get_distances
from ..prep.transforms import COCO_KEYPOINTS
COCO_KEYPOINTS = [
'nose', # 0
'left_eye', # 1
'right_eye', # 2
'left_ear', # 3
'right_ear', # 4
'left_shoulder', # 5
'right_shoulder', # 6
'left_elbow', # 7
'right_elbow', # 8
'left_wrist', # 9
'right_wrist', # 10
'left_hip', # 11
'right_hip', # 12
'left_knee', # 13
'right_knee', # 14
'left_ankle', # 15
'right_ankle', # 16
]
def joints_variance(joints, clusters, dic_ms): def joints_variance(joints, clusters, dic_ms):
@ -184,8 +165,8 @@ def variance_figures(dic_fin, clusters):
plt.title("Standard deviation of joints disparity") plt.title("Standard deviation of joints disparity")
yys_p = [el for _, el in dic_fin['pifpaf']['std_d'].items()] yys_p = [el for _, el in dic_fin['pifpaf']['std_d'].items()]
yys_m = [el for _, el in dic_fin['mask']['std_d'].items()] yys_m = [el for _, el in dic_fin['mask']['std_d'].items()]
yys_p_z = [el for _, el in dic_fin['pifpaf']['std_z'].items()] # yys_p_z = [el for _, el in dic_fin['pifpaf']['std_z'].items()]
yys_m_z = [el for _, el in dic_fin['mask']['std_z'].items()] # yys_m_z = [el for _, el in dic_fin['mask']['std_z'].items()]
plt.plot(xxs, yys_p, marker='s', label="PifPaf") plt.plot(xxs, yys_p, marker='s', label="PifPaf")
plt.plot(xxs, yys_m, marker='o', label="Mask R-CNN") plt.plot(xxs, yys_m, marker='o', label="Mask R-CNN")
# plt.plot(xxs, yys_p_z, marker='s', color='b', label="PifPaf (meters)") # plt.plot(xxs, yys_p_z, marker='s', color='b', label="PifPaf (meters)")

40
monoloco/eval/generate_kitti.py
View File

@ -13,10 +13,11 @@ import torch
from ..network import Loco from ..network import Loco
from ..network.process import preprocess_pifpaf from ..network.process import preprocess_pifpaf
from ..network.geom_baseline import geometric_coordinates from .geom_baseline import geometric_coordinates
from ..utils import get_keypoints, pixel_to_camera, factory_file, factory_basename, make_new_directory, get_category, \ from ..utils import get_keypoints, pixel_to_camera, factory_basename, make_new_directory, get_category, \
xyz_from_distance, read_and_rewrite xyz_from_distance, read_and_rewrite
from .stereo_baselines import baselines_association from .stereo_baselines import baselines_association
from ..prep import factory_file
from .reid_baseline import get_reid_features, ReID from .reid_baseline import get_reid_features, ReID
@ -93,7 +94,7 @@ class GenerateKitti:
make_new_directory(di) make_new_directory(di)
dir_out = {self.net: di} dir_out = {self.net: di}
for mode, names in self.baselines.items(): for _, names in self.baselines.items():
for name in names: for name in names:
di = os.path.join('data', 'kitti', name) di = os.path.join('data', 'kitti', name)
make_new_directory(di) make_new_directory(di)
@ -106,7 +107,7 @@ class GenerateKitti:
boxes, keypoints = preprocess_pifpaf(annotations, im_size=(1242, 374)) boxes, keypoints = preprocess_pifpaf(annotations, im_size=(1242, 374))
cat = get_category(keypoints, os.path.join(self.dir_byc, basename + '.json')) cat = get_category(keypoints, os.path.join(self.dir_byc, basename + '.json'))
if keypoints: if keypoints:
annotations_r, _, _ = factory_file(path_calib, self.dir_ann, basename, mode='right') annotations_r, _, _ = factory_file(path_calib, self.dir_ann, basename, ann_type='right')
_, keypoints_r = preprocess_pifpaf(annotations_r, im_size=(1242, 374)) _, keypoints_r = preprocess_pifpaf(annotations_r, im_size=(1242, 374))
if self.net == 'monstereo': if self.net == 'monstereo':
@ -121,7 +122,7 @@ class GenerateKitti:
# Save txt files # Save txt files
params = [kk, tt] params = [kk, tt]
path_txt = os.path.join(dir_out[self.net], basename + '.txt') path_txt = os.path.join(dir_out[self.net], basename + '.txt')
save_txts(path_txt, boxes, all_outputs[self.net], params, mode=self.net, cat=cat) save_txts(path_txt, boxes, all_outputs[self.net], params, net=self.net, cat=cat)
cnt_ann += len(boxes) cnt_ann += len(boxes)
cnt_file += 1 cnt_file += 1
@ -136,7 +137,7 @@ class GenerateKitti:
# monocular baselines # monocular baselines
for key in self.baselines['mono']: for key in self.baselines['mono']:
path_txt = {key: os.path.join(dir_out[key], basename + '.txt')} path_txt = {key: os.path.join(dir_out[key], basename + '.txt')}
save_txts(path_txt[key], boxes, all_outputs[key], params, mode=key, cat=cat) save_txts(path_txt[key], boxes, all_outputs[key], params, net=key, cat=cat)
# stereo baselines # stereo baselines
if self.baselines['stereo']: if self.baselines['stereo']:
@ -149,7 +150,7 @@ class GenerateKitti:
path_txt[key] = os.path.join(dir_out[key], basename + '.txt') path_txt[key] = os.path.join(dir_out[key], basename + '.txt')
save_txts(path_txt[key], all_inputs[key], all_outputs[key], params, save_txts(path_txt[key], all_inputs[key], all_outputs[key], params,
mode='baseline', net='baseline',
cat=cat) cat=cat)
print("\nSaved in {} txt {} annotations. Not found {} images".format(cnt_file, cnt_ann, cnt_no_file)) print("\nSaved in {} txt {} annotations. Not found {} images".format(cnt_file, cnt_ann, cnt_no_file))
@ -166,9 +167,9 @@ class GenerateKitti:
def _run_stereo_baselines(self, basename, boxes, keypoints, zzs, path_calib): def _run_stereo_baselines(self, basename, boxes, keypoints, zzs, path_calib):
annotations_r, _, _ = factory_file(path_calib, self.dir_ann, basename, mode='right') annotations_r, _, _ = factory_file(path_calib, self.dir_ann, basename, ann_type='right')
boxes_r, keypoints_r = preprocess_pifpaf(annotations_r, im_size=(1242, 374)) boxes_r, keypoints_r = preprocess_pifpaf(annotations_r, im_size=(1242, 374))
_, kk, tt = factory_file(path_calib, self.dir_ann, basename) _, kk, _ = factory_file(path_calib, self.dir_ann, basename)
uv_centers = get_keypoints(keypoints, mode='bottom') # Kitti uses the bottom center to calculate depth uv_centers = get_keypoints(keypoints, mode='bottom') # Kitti uses the bottom center to calculate depth
xy_centers = pixel_to_camera(uv_centers, kk, 1) xy_centers = pixel_to_camera(uv_centers, kk, 1)
@ -198,15 +199,15 @@ class GenerateKitti:
return dic_xyz return dic_xyz
def save_txts(path_txt, all_inputs, all_outputs, all_params, mode='monoloco', cat=None): def save_txts(path_txt, all_inputs, all_outputs, all_params, net='monoloco', cat=None):
assert mode in ('monoloco', 'monstereo', 'geometric', 'baseline', 'monoloco_pp') assert net in ('monoloco', 'monstereo', 'geometric', 'baseline', 'monoloco_pp')
if mode in ('monstereo', 'monoloco_pp'): if net in ('monstereo', 'monoloco_pp'):
xyzd, bis, epis, yaws, hs, ws, ls = all_outputs[:] xyzd, bis, epis, yaws, hs, ws, ls = all_outputs[:]
xyz = xyzd[:, 0:3] xyz = xyzd[:, 0:3]
tt = [0, 0, 0] tt = [0, 0, 0]
elif mode in ('monoloco', 'geometric'): elif net in ('monoloco', 'geometric'):
tt = [0, 0, 0] tt = [0, 0, 0]
dds, bis, epis, zzs_geom, xy_centers = all_outputs[:] dds, bis, epis, zzs_geom, xy_centers = all_outputs[:]
xyz = xyz_from_distance(dds, xy_centers) xyz = xyz_from_distance(dds, xy_centers)
@ -223,25 +224,18 @@ def save_txts(path_txt, all_inputs, all_outputs, all_params, mode='monoloco', ca
yy = float(xyz[idx][1]) - tt[1] yy = float(xyz[idx][1]) - tt[1]
zz = float(xyz[idx][2]) - tt[2] zz = float(xyz[idx][2]) - tt[2]
if mode == 'geometric': if net == 'geometric':
zz = zzs_geom[idx] zz = zzs_geom[idx]
cam_0 = [xx, yy, zz] cam_0 = [xx, yy, zz]
bi = float(bis[idx]) bi = float(bis[idx])
epi = float(epis[idx]) epi = float(epis[idx])
if mode in ('monstereo', 'monoloco_pp'): if net in ('monstereo', 'monoloco_pp'):
alpha, ry = float(yaws[0][idx]), float(yaws[1][idx]) alpha, ry = float(yaws[0][idx]), float(yaws[1][idx])
hwl = [float(hs[idx]), float(ws[idx]), float(ls[idx])] hwl = [float(hs[idx]), float(ws[idx]), float(ls[idx])]
conf_scale = 0.035 # scale to obtain (approximately) same recall at evaluation
else: else:
alpha, ry, hwl = -10., -10., [0, 0, 0] alpha, ry, hwl = -10., -10., [0, 0, 0]
# Set the scale to obtain (approximately) same recall at evaluation
if mode == 'monstereo':
conf_scale = 0.03
elif mode == 'monoloco_pp':
conf_scale = 0.033
# conf_scale = 0.035 # nuScenes for having same recall
else:
conf_scale = 0.05 conf_scale = 0.05
conf = conf_scale * (uv_box[-1]) / (bi / math.sqrt(xx ** 2 + yy ** 2 + zz ** 2)) conf = conf_scale * (uv_box[-1]) / (bi / math.sqrt(xx ** 2 + yy ** 2 + zz ** 2))

43
monoloco/eval/geom_baseline.py
View File

@ -1,17 +1,34 @@
import json import json
import logging
import math import math
from collections import defaultdict from collections import defaultdict
import numpy as np import numpy as np
from ..utils import pixel_to_camera, get_keypoints from monoloco.utils import pixel_to_camera, get_keypoints
AVERAGE_Y = 0.48 AVERAGE_Y = 0.48
CLUSTERS = ['10', '20', '30', 'all'] CLUSTERS = ['10', '20', '30', 'all']
def geometric_coordinates(keypoints, kk, average_y=0.48):
""" Evaluate geometric depths for a set of keypoints"""
zzs_geom = []
uv_shoulders = get_keypoints(keypoints, mode='shoulder')
uv_hips = get_keypoints(keypoints, mode='hip')
uv_centers = get_keypoints(keypoints, mode='center')
xy_shoulders = pixel_to_camera(uv_shoulders, kk, 1)
xy_hips = pixel_to_camera(uv_hips, kk, 1)
xy_centers = pixel_to_camera(uv_centers, kk, 1)
for idx, xy_shoulder in enumerate(xy_shoulders):
zz = compute_depth(xy_shoulder, xy_hips[idx], average_y)
zzs_geom.append(zz)
return zzs_geom, xy_centers
def geometric_baseline(joints): def geometric_baseline(joints):
""" """
List of json files --> 2 lists with mean and std for each segment and the total count of instances List of json files --> 2 lists with mean and std for each segment and the total count of instances
@ -28,8 +45,6 @@ def geometric_baseline(joints):
'right_ankle'] 'right_ankle']
""" """
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
cnt_tot = 0 cnt_tot = 0
dic_dist = defaultdict(lambda: defaultdict(list)) dic_dist = defaultdict(lambda: defaultdict(list))
@ -48,13 +63,13 @@ def geometric_baseline(joints):
errors = calculate_error(dic_dist['error']) errors = calculate_error(dic_dist['error'])
# Show results # Show results
logger.info("Computed distance of {} annotations".format(cnt_tot)) print("Computed distance of {} annotations".format(cnt_tot))
for key in dic_h_means: for key in dic_h_means:
logger.info("Average height of segment {} is {:.2f} with a std of {:.2f}". print("Average height of segment {} is {:.2f} with a std of {:.2f}".
format(key, dic_h_means[key], dic_h_stds[key])) format(key, dic_h_means[key], dic_h_stds[key]))
for clst in CLUSTERS: for clst in CLUSTERS:
logger.info("Average error over the val set for clst {}: {:.2f}".format(clst, errors[clst])) print("Average error over the val set for clst {}: {:.2f}".format(clst, errors[clst]))
logger.info("Joints used: {}".format(joints)) print("Joints used: {}".format(joints))
def update_distances(dic_fin, dic_dist, phase, average_y): def update_distances(dic_fin, dic_dist, phase, average_y):
@ -78,9 +93,9 @@ def update_distances(dic_fin, dic_dist, phase, average_y):
dy_met = abs(float((dic_xyz['hip'][0][1] - dic_xyz['shoulder'][0][1]))) dy_met = abs(float((dic_xyz['hip'][0][1] - dic_xyz['shoulder'][0][1])))
# Estimate distance for a single annotation # Estimate distance for a single annotation
z_met_real = compute_distance(dic_xyz_norm['shoulder'][0], dic_xyz_norm['hip'][0], average_y, z_met_real = compute_depth(dic_xyz_norm['shoulder'][0], dic_xyz_norm['hip'][0], average_y,
mode='real', dy_met=dy_met) mode='real', dy_met=dy_met)
z_met_approx = compute_distance(dic_xyz_norm['shoulder'][0], dic_xyz_norm['hip'][0], average_y, mode='average') z_met_approx = compute_depth(dic_xyz_norm['shoulder'][0], dic_xyz_norm['hip'][0], average_y, mode='average')
# Compute distance with respect to the center of the 3D bounding box # Compute distance with respect to the center of the 3D bounding box
d_real = math.sqrt(z_met_real ** 2 + dic_fin['boxes_3d'][idx][0] ** 2 + dic_fin['boxes_3d'][idx][1] ** 2) d_real = math.sqrt(z_met_real ** 2 + dic_fin['boxes_3d'][idx][0] ** 2 + dic_fin['boxes_3d'][idx][1] ** 2)
@ -94,9 +109,9 @@ def update_distances(dic_fin, dic_dist, phase, average_y):
return cnt return cnt
def compute_distance(xyz_norm_1, xyz_norm_2, average_y, mode='average', dy_met=0): def compute_depth(xyz_norm_1, xyz_norm_2, average_y, mode='average', dy_met=0):
""" """
Compute distance Z of a mask annotation (solving a linear system) for 2 possible cases: Compute depth Z of a mask annotation (solving a linear system) for 2 possible cases:
1. knowing specific height of the annotation (head-ankle) dy_met 1. knowing specific height of the annotation (head-ankle) dy_met
2. using mean height of people (average_y) 2. using mean height of people (average_y)
""" """

2
monoloco/eval/reid_baseline.py
View File

@ -27,7 +27,7 @@ def get_reid_features(reid_net, boxes, boxes_r, path_image, path_image_r):
return features.cpu(), features_r.cpu() return features.cpu(), features_r.cpu()
class ReID(object): class ReID:
def __init__(self, weights_path, device, num_classes=751, height=256, width=128): def __init__(self, weights_path, device, num_classes=751, height=256, width=128):
super().__init__() super().__init__()
torch.manual_seed(1) torch.manual_seed(1)

2
monoloco/eval/stereo_baselines.py
View File

@ -22,7 +22,7 @@ def baselines_association(baselines, zzs, keypoints, keypoints_right, reid_featu
cnt_stereo['max'] = min(keypoints.shape[0], keypoints_r.shape[0]) # pylint: disable=E1136 cnt_stereo['max'] = min(keypoints.shape[0], keypoints_r.shape[0]) # pylint: disable=E1136
# Filter joints disparity and calculate avg disparity # Filter joints disparity and calculate avg disparity
avg_disparities, disparities_x, disparities_y = mask_joint_disparity(keypoints, keypoints_r) avg_disparities, _, _ = mask_joint_disparity(keypoints, keypoints_r)
# Iterate over each left pose # Iterate over each left pose
for key in baselines: for key in baselines:

2
monoloco/network/architectures.py
View File

@ -3,7 +3,7 @@ import torch
import torch.nn as nn import torch.nn as nn
class MonStereoModel(nn.Module): class LocoModel(nn.Module):
def __init__(self, input_size, output_size=2, linear_size=512, p_dropout=0.2, num_stage=3, device='cuda'): def __init__(self, input_size, output_size=2, linear_size=512, p_dropout=0.2, num_stage=3, device='cuda'):
super().__init__() super().__init__()

213
monoloco/network/geom_baseline.py
View File

@ -1,213 +0,0 @@
import json
import logging
import math
from collections import defaultdict
import numpy as np
from monoloco.utils import pixel_to_camera, get_keypoints
AVERAGE_Y = 0.48
CLUSTERS = ['10', '20', '30', 'all']
def geometric_coordinates(keypoints, kk, average_y=0.48):
""" Evaluate geometric depths for a set of keypoints"""
zzs_geom = []
uv_shoulders = get_keypoints(keypoints, mode='shoulder')
uv_hips = get_keypoints(keypoints, mode='hip')
uv_centers = get_keypoints(keypoints, mode='center')
xy_shoulders = pixel_to_camera(uv_shoulders, kk, 1)
xy_hips = pixel_to_camera(uv_hips, kk, 1)
xy_centers = pixel_to_camera(uv_centers, kk, 1)
for idx, xy_shoulder in enumerate(xy_shoulders):
zz = compute_depth(xy_shoulders[idx], xy_hips[idx], average_y)
zzs_geom.append(zz)
return zzs_geom, xy_centers
def geometric_baseline(joints):
"""
List of json files --> 2 lists with mean and std for each segment and the total count of instances
For each annotation:
1. From gt boxes calculate the height (deltaY) for the segments head, shoulder, hip, ankle
2. From mask boxes calculate distance of people using average height of people and real pixel height
For left-right ambiguities we chose always the average of the joints
The joints are mapped from 0 to 16 in the following order:
['nose', 'left_eye', 'right_eye', 'left_ear', 'right_ear', 'left_shoulder', 'right_shoulder', 'left_elbow',
'right_elbow', 'left_wrist', 'right_wrist', 'left_hip', 'right_hip', 'left_knee', 'right_knee', 'left_ankle',
'right_ankle']
"""
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
cnt_tot = 0
dic_dist = defaultdict(lambda: defaultdict(list))
# Access the joints file
with open(joints, 'r') as ff:
dic_joints = json.load(ff)
# Calculate distances for all the instances in the joints dictionary
for phase in ['train', 'val']:
cnt = update_distances(dic_joints[phase], dic_dist, phase, AVERAGE_Y)
cnt_tot += cnt
# Calculate mean and std of each segment
dic_h_means = calculate_heights(dic_dist['heights'], mode='mean')
dic_h_stds = calculate_heights(dic_dist['heights'], mode='std')
errors = calculate_error(dic_dist['error'])
# Show results
logger.info("Computed distance of {} annotations".format(cnt_tot))
for key in dic_h_means:
logger.info("Average height of segment {} is {:.2f} with a std of {:.2f}".
format(key, dic_h_means[key], dic_h_stds[key]))
for clst in CLUSTERS:
logger.info("Average error over the val set for clst {}: {:.2f}".format(clst, errors[clst]))
logger.info("Joints used: {}".format(joints))
def update_distances(dic_fin, dic_dist, phase, average_y):
# Loop over each annotation in the json file corresponding to the image
cnt = 0
for idx, kps in enumerate(dic_fin['kps']):
# Extract pixel coordinates of head, shoulder, hip, ankle and and save them
dic_uv = {mode: get_keypoints(kps, mode) for mode in ['head', 'shoulder', 'hip', 'ankle']}
# Convert segments from pixel coordinate to camera coordinate
kk = dic_fin['K'][idx]
z_met = dic_fin['boxes_3d'][idx][2]
# Create a dict with all annotations in meters
dic_xyz = {key: pixel_to_camera(dic_uv[key], kk, z_met) for key in dic_uv}
dic_xyz_norm = {key: pixel_to_camera(dic_uv[key], kk, 1) for key in dic_uv}
# Compute real height
dy_met = abs(float((dic_xyz['hip'][0][1] - dic_xyz['shoulder'][0][1])))
# Estimate distance for a single annotation
z_met_real = compute_depth(dic_xyz_norm['shoulder'][0], dic_xyz_norm['hip'][0], average_y,
mode='real', dy_met=dy_met)
z_met_approx = compute_depth(dic_xyz_norm['shoulder'][0], dic_xyz_norm['hip'][0], average_y, mode='average')
# Compute distance with respect to the center of the 3D bounding box
d_real = math.sqrt(z_met_real ** 2 + dic_fin['boxes_3d'][idx][0] ** 2 + dic_fin['boxes_3d'][idx][1] ** 2)
d_approx = math.sqrt(z_met_approx ** 2 +
dic_fin['boxes_3d'][idx][0] ** 2 + dic_fin['boxes_3d'][idx][1] ** 2)
# Update the dictionary with distance and heights metrics
dic_dist = update_dic_dist(dic_dist, dic_xyz, d_real, d_approx, phase)
cnt += 1
return cnt
def compute_depth(xyz_norm_1, xyz_norm_2, average_y, mode='average', dy_met=0):
"""
Compute depth Z of a mask annotation (solving a linear system) for 2 possible cases:
1. knowing specific height of the annotation (head-ankle) dy_met
2. using mean height of people (average_y)
"""
assert mode in ('average', 'real')
x1 = float(xyz_norm_1[0])
y1 = float(xyz_norm_1[1])
x2 = float(xyz_norm_2[0])
y2 = float(xyz_norm_2[1])
xx = (x1 + x2) / 2
# Choose if solving for provided height or average one.
if mode == 'average':
cc = - average_y # Y axis goes down
else:
cc = -dy_met
# Solving the linear system Ax = b
matrix = np.array([[y1, 0, -xx],
[0, -y1, 1],
[y2, 0, -xx],
[0, -y2, 1]])
bb = np.array([cc * xx, -cc, 0, 0]).reshape(4, 1)
xx = np.linalg.lstsq(matrix, bb, rcond=None)
z_met = abs(np.float(xx[0][1])) # Abs take into account specularity behind the observer
return z_met
def update_dic_dist(dic_dist, dic_xyz, d_real, d_approx, phase):
""" For every annotation in a single image, update the final dictionary"""
# Update the dict with heights metric
if phase == 'train':
dic_dist['heights']['head'].append(float(dic_xyz['head'][0][1]))
dic_dist['heights']['shoulder'].append(float(dic_xyz['shoulder'][0][1]))
dic_dist['heights']['hip'].append(float(dic_xyz['hip'][0][1]))
dic_dist['heights']['ankle'].append(float(dic_xyz['ankle'][0][1]))
# Update the dict with distance metrics for the test phase
if phase == 'val':
error = abs(d_real - d_approx)
if d_real <= 10:
dic_dist['error']['10'].append(error)
elif d_real <= 20:
dic_dist['error']['20'].append(error)
elif d_real <= 30:
dic_dist['error']['30'].append(error)
else:
dic_dist['error']['>30'].append(error)
dic_dist['error']['all'].append(error)
return dic_dist
def calculate_heights(heights, mode):
"""
Compute statistics of heights based on the distance
"""
assert mode in ('mean', 'std', 'max')
heights_fin = {}
head_shoulder = np.array(heights['shoulder']) - np.array(heights['head'])
shoulder_hip = np.array(heights['hip']) - np.array(heights['shoulder'])
hip_ankle = np.array(heights['ankle']) - np.array(heights['hip'])
if mode == 'mean':
heights_fin['head_shoulder'] = np.float(np.mean(head_shoulder)) * 100
heights_fin['shoulder_hip'] = np.float(np.mean(shoulder_hip)) * 100
heights_fin['hip_ankle'] = np.float(np.mean(hip_ankle)) * 100
elif mode == 'std':
heights_fin['head_shoulder'] = np.float(np.std(head_shoulder)) * 100
heights_fin['shoulder_hip'] = np.float(np.std(shoulder_hip)) * 100
heights_fin['hip_ankle'] = np.float(np.std(hip_ankle)) * 100
elif mode == 'max':
heights_fin['head_shoulder'] = np.float(np.max(head_shoulder)) * 100
heights_fin['shoulder_hip'] = np.float(np.max(shoulder_hip)) * 100
heights_fin['hip_ankle'] = np.float(np.max(hip_ankle)) * 100
return heights_fin
def calculate_error(dic_errors):
"""
Compute statistics of distances based on the distance
"""
errors = {}
for clst in dic_errors:
errors[clst] = np.float(np.mean(np.array(dic_errors[clst])))
return errors

141
monoloco/network/losses.py
View File

@ -1,141 +0,0 @@
import math
import torch
import numpy as np
import matplotlib.pyplot as plt
class CustomL1Loss(torch.nn.Module):
"""
L1 loss with more weight to errors at a shorter distance
It inherits from nn.module so it supports backward
"""
def __init__(self, dic_norm, device, beta=1):
super().__init__()
self.dic_norm = dic_norm
self.device = device
self.beta = beta
@staticmethod
def compute_weights(xx, beta=1):
"""
Return the appropriate weight depending on the distance and the hyperparameter chosen
alpha = 1 refers to the curve of A Photogrammetric Approach for Real-time...
It is made for unnormalized outputs (to be more understandable)
From 70 meters on every value is weighted the same (0.1**beta)
Alpha is optional value from Focal loss. Yet to be analyzed
"""
# alpha = np.maximum(1, 10 ** (beta - 1))
alpha = 1
ww = np.maximum(0.1, 1 - xx / 78)**beta
return alpha * ww
def print_loss(self):
xx = np.linspace(0, 80, 100)
y1 = self.compute_weights(xx, beta=1)
y2 = self.compute_weights(xx, beta=2)
y3 = self.compute_weights(xx, beta=3)
plt.plot(xx, y1)
plt.plot(xx, y2)
plt.plot(xx, y3)
plt.xlabel("Distance [m]")
plt.ylabel("Loss function Weight")
plt.legend(("Beta = 1", "Beta = 2", "Beta = 3"))
plt.show()
def forward(self, output, target):
unnormalized_output = output.cpu().detach().numpy() * self.dic_norm['std']['Y'] + self.dic_norm['mean']['Y']
weights_np = self.compute_weights(unnormalized_output, self.beta)
weights = torch.from_numpy(weights_np).float().to(self.device) # To make weights in the same cuda device
losses = torch.abs(output - target) * weights
loss = losses.mean() # Mean over the batch
return loss
class LaplacianLoss(torch.nn.Module):
"""1D Gaussian with std depending on the absolute distance
"""
def __init__(self, size_average=True, reduce=True, evaluate=False):
super().__init__()
self.size_average = size_average
self.reduce = reduce
self.evaluate = evaluate
def laplacian_1d(self, mu_si, xx):
"""
1D Gaussian Loss. f(x | mu, sigma). The network outputs mu and sigma. X is the ground truth distance.
This supports backward().
Inspired by
https://github.com/naba89/RNN-Handwriting-Generation-Pytorch/blob/master/loss_functions.py
"""
mu, si = mu_si[:, 0:1], mu_si[:, 1:2]
# norm = xx - mu
norm = 1 - mu / xx # Relative
term_a = torch.abs(norm) * torch.exp(-si)
term_b = si
norm_bi = (np.mean(np.abs(norm.cpu().detach().numpy())), np.mean(torch.exp(si).cpu().detach().numpy()))
if self.evaluate:
return norm_bi
return term_a + term_b
def forward(self, outputs, targets):
values = self.laplacian_1d(outputs, targets)
if not self.reduce or self.evaluate:
return values
if self.size_average:
mean_values = torch.mean(values)
return mean_values
return torch.sum(values)
class GaussianLoss(torch.nn.Module):
"""1D Gaussian with std depending on the absolute distance
"""
def __init__(self, device, size_average=True, reduce=True, evaluate=False):
super().__init__()
self.size_average = size_average
self.reduce = reduce
self.evaluate = evaluate
self.device = device
def gaussian_1d(self, mu_si, xx):
"""
1D Gaussian Loss. f(x | mu, sigma). The network outputs mu and sigma. X is the ground truth distance.
This supports backward().
Inspired by
https://github.com/naba89/RNN-Handwriting-Generation-Pytorch/blob/master/loss_functions.py
"""
mu, si = mu_si[:, 0:1], mu_si[:, 1:2]
min_si = torch.ones(si.size()).cuda(self.device) * 0.1
si = torch.max(min_si, si)
norm = xx - mu
term_a = (norm / si)**2 / 2
term_b = torch.log(si * math.sqrt(2 * math.pi))
norm_si = (np.mean(np.abs(norm.cpu().detach().numpy())), np.mean(si.cpu().detach().numpy()))
if self.evaluate:
return norm_si
return term_a + term_b
def forward(self, outputs, targets):
values = self.gaussian_1d(outputs, targets)
if not self.reduce or self.evaluate:
return values
if self.size_average:
mean_values = torch.mean(values)
return mean_values
return torch.sum(values)

15
monoloco/network/net.py
View File

@ -9,13 +9,15 @@ import math
import logging import logging
from collections import defaultdict from collections import defaultdict
import numpy as np
import torch import torch
from ..utils import get_iou_matches, reorder_matches, get_keypoints, pixel_to_camera, xyz_from_distance from ..utils import get_iou_matches, reorder_matches, get_keypoints, pixel_to_camera, xyz_from_distance, \
mask_joint_disparity
from .process import preprocess_monstereo, preprocess_monoloco, extract_outputs, extract_outputs_mono,\ from .process import preprocess_monstereo, preprocess_monoloco, extract_outputs, extract_outputs_mono,\
filter_outputs, cluster_outputs, unnormalize_bi filter_outputs, cluster_outputs, unnormalize_bi, laplace_sampling
from ..activity import social_interactions from ..activity import social_interactions
from .architectures import MonolocoModel, MonStereoModel from .architectures import MonolocoModel, LocoModel
class Loco: class Loco:
@ -69,7 +71,7 @@ class Loco:
self.model = MonolocoModel(p_dropout=p_dropout, input_size=input_size, linear_size=linear_size, self.model = MonolocoModel(p_dropout=p_dropout, input_size=input_size, linear_size=linear_size,
output_size=output_size) output_size=output_size)
else: else:
self.model = MonStereoModel(p_dropout=p_dropout, input_size=input_size, output_size=output_size, self.model = LocoModel(p_dropout=p_dropout, input_size=input_size, output_size=output_size,
linear_size=linear_size, device=self.device) linear_size=linear_size, device=self.device)
self.model.load_state_dict(torch.load(model_path, map_location=lambda storage, loc: storage)) self.model.load_state_dict(torch.load(model_path, map_location=lambda storage, loc: storage))
@ -116,7 +118,7 @@ class Loco:
outputs = self.model(inputs) outputs = self.model(inputs)
outputs = cluster_outputs(outputs, keypoints_r.shape[0]) outputs = cluster_outputs(outputs, keypoints_r.shape[0])
outputs_fin, mask = filter_outputs(outputs) outputs_fin, _ = filter_outputs(outputs)
dic_out = extract_outputs(outputs_fin) dic_out = extract_outputs(outputs_fin)
# For Median baseline # For Median baseline
@ -136,7 +138,6 @@ class Loco:
Apply dropout at test time to obtain combined aleatoric + epistemic uncertainty Apply dropout at test time to obtain combined aleatoric + epistemic uncertainty
""" """
assert self.net in ('monoloco', 'monoloco_p', 'monoloco_pp'), "Not supported for MonStereo" assert self.net in ('monoloco', 'monoloco_p', 'monoloco_pp'), "Not supported for MonStereo"
from .process import laplace_sampling
self.model.dropout.training = True # Manually reactivate dropout in eval self.model.dropout.training = True # Manually reactivate dropout in eval
total_outputs = torch.empty((0, inputs.size()[0])).to(self.device) total_outputs = torch.empty((0, inputs.size()[0])).to(self.device)
@ -271,8 +272,6 @@ def median_disparity(dic_out, keypoints, keypoints_r, mask):
Filters are applied to masks nan joints and remove outlier disparities with iqr Filters are applied to masks nan joints and remove outlier disparities with iqr
The mask input is used to filter the all-vs-all approach The mask input is used to filter the all-vs-all approach
""" """
import numpy as np
from ..utils import mask_joint_disparity
keypoints = keypoints.cpu().numpy() keypoints = keypoints.cpu().numpy()
keypoints_r = keypoints_r.cpu().numpy() keypoints_r = keypoints_r.cpu().numpy()

5
monoloco/network/process.py
View File

@ -7,7 +7,7 @@ import numpy as np
import torch import torch
import torchvision import torchvision
from ..utils import get_keypoints, pixel_to_camera, to_cartesian, back_correct_angles from ..utils import get_keypoints, pixel_to_camera, to_cartesian, back_correct_angles, open_annotations
logging.basicConfig(level=logging.INFO) logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@ -30,7 +30,7 @@ def preprocess_monstereo(keypoints, keypoints_r, kk):
inputs_r = preprocess_monoloco(keypoints_r, kk) inputs_r = preprocess_monoloco(keypoints_r, kk)
inputs = torch.empty((0, 68)).to(inputs_l.device) inputs = torch.empty((0, 68)).to(inputs_l.device)
for idx, inp_l in enumerate(inputs_l.split(1)): for inp_l in inputs_l.split(1):
clst = 0 clst = 0
# inp_l = torch.cat((inp_l, cat[:, idx:idx+1]), dim=1) # inp_l = torch.cat((inp_l, cat[:, idx:idx+1]), dim=1)
for idx_r, inp_r in enumerate(inputs_r.split(1)): for idx_r, inp_r in enumerate(inputs_r.split(1)):
@ -135,7 +135,6 @@ def preprocess_mask(dir_ann, basename, mode='left'):
elif mode == 'right': elif mode == 'right':
path_ann = os.path.join(dir_ann + '_right', basename + '.json') path_ann = os.path.join(dir_ann + '_right', basename + '.json')
from ..utils import open_annotations
dic = open_annotations(path_ann) dic = open_annotations(path_ann)
if isinstance(dic, list): if isinstance(dic, list):
return [], [] return [], []

41
monoloco/predict.py
View File

@ -1,12 +1,15 @@
# pylint: disable=too-many-statements, too-many-branches, undefined-loop-variable # pylint: disable=too-many-statements, too-many-branches, undefined-loop-variable
""" """
Adapted from https://github.com/vita-epfl/openpifpaf/blob/master/openpifpaf/predict.py Adapted from https://github.com/openpifpaf/openpifpaf/blob/main/openpifpaf/predict.py,
which is: 'Copyright 2019-2021 by Sven Kreiss and contributors. All rights reserved.'
and licensed under GNU AGPLv3
""" """
import os import os
import glob import glob
import json import json
import copy
import logging import logging
from collections import defaultdict from collections import defaultdict
@ -17,7 +20,11 @@ import openpifpaf
import openpifpaf.datasets as datasets import openpifpaf.datasets as datasets
from openpifpaf.predict import processor_factory, preprocess_factory from openpifpaf.predict import processor_factory, preprocess_factory
from openpifpaf import decoder, network, visualizer, show, logger from openpifpaf import decoder, network, visualizer, show, logger
try:
import gdown
DOWNLOAD = copy.copy(gdown.download)
except ImportError:
DOWNLOAD = None
from .visuals.printer import Printer from .visuals.printer import Printer
from .network import Loco from .network import Loco
from .network.process import factory_for_gt, preprocess_pifpaf from .network.process import factory_for_gt, preprocess_pifpaf
@ -46,12 +53,15 @@ def get_torch_checkpoints_dir():
def download_checkpoints(args): def download_checkpoints(args):
torch_dir = get_torch_checkpoints_dir() torch_dir = get_torch_checkpoints_dir()
pifpaf_model = os.path.join(torch_dir, 'shufflenetv2k30-201104-224654-cocokp-d75ed641.pkl') if args.checkpoint is None:
pifpaf_model = os.path.join(torch_dir, 'shufflenetv2k30-201104-224654-cocokp-d75ed641.pkl')
else:
pifpaf_model = args.checkpoint
dic_models = {'keypoints': pifpaf_model} dic_models = {'keypoints': pifpaf_model}
if not os.path.exists(pifpaf_model): if not os.path.exists(pifpaf_model):
import gdown assert DOWNLOAD is not None, "pip install gdown to download pifpaf model, or pass it as --checkpoint"
LOG.info('Downloading OpenPifPaf model in %s', torch_dir) LOG.info('Downloading OpenPifPaf model in %s', torch_dir)
gdown.download(OPENPIFPAF_MODEL, pifpaf_model, quiet=False) DOWNLOAD(OPENPIFPAF_MODEL, pifpaf_model, quiet=False)
if args.mode == 'keypoints': if args.mode == 'keypoints':
return dic_models return dic_models
@ -73,9 +83,9 @@ def download_checkpoints(args):
model = os.path.join(torch_dir, name) model = os.path.join(torch_dir, name)
dic_models[args.mode] = model dic_models[args.mode] = model
if not os.path.exists(model): if not os.path.exists(model):
import gdown assert DOWNLOAD is not None, "pip install gdown to download monoloco model, or pass it as --model"
LOG.info('Downloading model in %s', torch_dir) LOG.info('Downloading model in %s', torch_dir)
gdown.download(path, model, quiet=False) DOWNLOAD(path, model, quiet=False)
return dic_models return dic_models
@ -209,7 +219,7 @@ def predict(args):
else: else:
LOG.info("Prediction with MonStereo") LOG.info("Prediction with MonStereo")
boxes_r, keypoints_r = preprocess_pifpaf(pifpaf_outs['right'], im_size) _, keypoints_r = preprocess_pifpaf(pifpaf_outs['right'], im_size)
dic_out = net.forward(keypoints, kk, keypoints_r=keypoints_r) dic_out = net.forward(keypoints, kk, keypoints_r=keypoints_r)
dic_out = net.post_process(dic_out, boxes, keypoints, kk, dic_gt) dic_out = net.post_process(dic_out, boxes, keypoints, kk, dic_gt)
@ -229,15 +239,19 @@ def factory_outputs(args, pifpaf_outs, dic_out, output_path, kk=None):
# Verify conflicting options # Verify conflicting options
if any((xx in args.output_types for xx in ['front', 'bird', 'multi'])): if any((xx in args.output_types for xx in ['front', 'bird', 'multi'])):
assert args.mode != 'keypoints', "for keypoints please use pifpaf original arguments" assert args.mode != 'keypoints', "for keypoints please use pifpaf original arguments"
if args.social_distance: else:
assert args.mode == 'mono', "Social distancing only works with monocular network" assert 'json' in args.output_types or args.mode == 'keypoints', \
"No output saved, please select one among front, bird, multi, json, or pifpaf arguments"
if args.social_distance:
assert args.mode == 'mono', "Social distancing only works with monocular network"
if args.mode == 'keypoints': if args.mode == 'keypoints':
annotation_painter = openpifpaf.show.AnnotationPainter() annotation_painter = openpifpaf.show.AnnotationPainter()
with openpifpaf.show.image_canvas(pifpaf_outs['image'], output_path) as ax: with openpifpaf.show.image_canvas(pifpaf_outs['image'], output_path) as ax:
annotation_painter.annotations(ax, pifpaf_outs['pred']) annotation_painter.annotations(ax, pifpaf_outs['pred'])
return
elif any((xx in args.output_types for xx in ['front', 'bird', 'multi'])): if any((xx in args.output_types for xx in ['front', 'bird', 'multi'])):
LOG.info(output_path) LOG.info(output_path)
if args.social_distance: if args.social_distance:
show_social(args, pifpaf_outs['image'], output_path, pifpaf_outs['left'], dic_out) show_social(args, pifpaf_outs['image'], output_path, pifpaf_outs['left'], dic_out)
@ -246,9 +260,6 @@ def factory_outputs(args, pifpaf_outs, dic_out, output_path, kk=None):
figures, axes = printer.factory_axes(dic_out) figures, axes = printer.factory_axes(dic_out)
printer.draw(figures, axes, pifpaf_outs['image']) printer.draw(figures, axes, pifpaf_outs['image'])
elif 'json' in args.output_types: if 'json' in args.output_types:
with open(os.path.join(output_path + '.monoloco.json'), 'w') as ff: with open(os.path.join(output_path + '.monoloco.json'), 'w') as ff:
json.dump(dic_out, ff) json.dump(dic_out, ff)
else:
LOG.info("No output saved, please select one among front, bird, multi, or pifpaf options")

2
monoloco/prep/__init__.py
View File

@ -0,0 +1,2 @@
from .preprocess_kitti import parse_ground_truth, factory_file

350
monoloco/prep/prep_kitti.py
View File

@ -1,350 +0,0 @@
# pylint: disable=too-many-statements, too-many-branches, too-many-nested-blocks
"""Preprocess annotations with KITTI ground-truth"""
import os
import glob
import copy
import logging
from collections import defaultdict
import json
import datetime
from PIL import Image
import torch
import cv2
from ..utils import split_training, parse_ground_truth, get_iou_matches, append_cluster, factory_file, \
extract_stereo_matches, get_category, normalize_hwl, make_new_directory
from ..network.process import preprocess_pifpaf, preprocess_monoloco
from .transforms import flip_inputs, flip_labels, height_augmentation
class PreprocessKitti:
"""Prepare arrays with same format as nuScenes preprocessing but using ground truth txt files"""
dir_gt = os.path.join('data', 'kitti', 'gt')
dir_images = '/data/lorenzo-data/kitti/original_images/training/image_2'
dir_byc_l = '/data/lorenzo-data/kitti/object_detection/left'
# SOCIAL DISTANCING PARAMETERS
THRESHOLD_DIST = 2 # Threshold to check distance of people
RADII = (0.3, 0.5, 1) # expected radii of the o-space
SOCIAL_DISTANCE = True
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
dic_jo = {'train': dict(X=[], Y=[], names=[], kps=[], K=[],
clst=defaultdict(lambda: defaultdict(list))),
'val': dict(X=[], Y=[], names=[], kps=[], K=[],
clst=defaultdict(lambda: defaultdict(list))),
'test': dict(X=[], Y=[], names=[], kps=[], K=[],
clst=defaultdict(lambda: defaultdict(list)))}
dic_names = defaultdict(lambda: defaultdict(list))
dic_std = defaultdict(lambda: defaultdict(list))
def __init__(self, dir_ann, mode='mono', iou_min=0.3):
self.dir_ann = dir_ann
self.iou_min = iou_min
self.mode = mode
assert self.mode in ('mono', 'stereo'), "modality not recognized"
self.names_gt = tuple(os.listdir(self.dir_gt))
self.dir_kk = os.path.join('data', 'kitti', 'calib')
self.list_gt = glob.glob(self.dir_gt + '/*.txt')
assert os.path.exists(self.dir_gt), "Ground truth dir does not exist"
assert os.path.exists(self.dir_ann), "Annotation dir does not exist"
now = datetime.datetime.now()
now_time = now.strftime("%Y%m%d-%H%M")[2:]
dir_out = os.path.join('data', 'arrays')
self.path_joints = os.path.join(dir_out, 'joints-kitti-' + now_time + '.json')
self.path_names = os.path.join(dir_out, 'names-kitti-' + now_time + '.json')
path_train = os.path.join('splits', 'kitti_train.txt')
path_val = os.path.join('splits', 'kitti_val.txt')
self.set_train, self.set_val = split_training(self.names_gt, path_train, path_val)
def run(self):
cnt_match_l, cnt_match_r, cnt_pair, cnt_pair_tot, cnt_extra_pair, cnt_files, cnt_files_ped, cnt_fnf, \
cnt_tot, cnt_ambiguous, cnt_cyclist = 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
cnt_mono = {'train': 0, 'val': 0, 'test': 0}
cnt_gt = cnt_mono.copy()
cnt_stereo = cnt_mono.copy()
correct_ped, correct_byc, wrong_ped, wrong_byc = 0, 0, 0, 0
cnt_30, cnt_less_30 = 0, 0
# self.names_gt = ('002282.txt',)
for name in self.names_gt:
path_gt = os.path.join(self.dir_gt, name)
basename, _ = os.path.splitext(name)
path_im = os.path.join(self.dir_images, basename + '.png')
phase, flag = self._factory_phase(name)
if flag:
cnt_fnf += 1
continue
if phase == 'train':
min_conf = 0
category = 'all'
else: # Remove for original results
min_conf = 0.1
category = 'pedestrian'
# Extract ground truth
boxes_gt, ys, _, _ = parse_ground_truth(path_gt, # pylint: disable=unbalanced-tuple-unpacking
category=category,
spherical=True)
cnt_gt[phase] += len(boxes_gt)
cnt_files += 1
cnt_files_ped += min(len(boxes_gt), 1) # if no boxes 0 else 1
# Extract keypoints
path_calib = os.path.join(self.dir_kk, basename + '.txt')
annotations, kk, tt = factory_file(path_calib, self.dir_ann, basename)
self.dic_names[basename + '.png']['boxes'] = copy.deepcopy(boxes_gt)
self.dic_names[basename + '.png']['ys'] = copy.deepcopy(ys)
self.dic_names[basename + '.png']['K'] = copy.deepcopy(kk)
# Check image size
with Image.open(path_im) as im:
width, height = im.size
boxes, keypoints = preprocess_pifpaf(annotations, im_size=(width, height), min_conf=min_conf)
if keypoints:
annotations_r, kk_r, tt_r = factory_file(path_calib, self.dir_ann, basename, mode='right')
boxes_r, keypoints_r = preprocess_pifpaf(annotations_r, im_size=(width, height), min_conf=min_conf)
cat = get_category(keypoints, os.path.join(self.dir_byc_l, basename + '.json'))
if not keypoints_r: # Case of no detection
all_boxes_gt, all_ys = [boxes_gt], [ys]
boxes_r, keypoints_r = boxes[0:1].copy(), keypoints[0:1].copy()
all_boxes, all_keypoints = [boxes], [keypoints]
all_keypoints_r = [keypoints_r]
else:
# Horizontal Flipping for training
if phase == 'train':
# GT)
boxes_gt_flip, ys_flip = flip_labels(boxes_gt, ys, im_w=width)
# New left
boxes_flip = flip_inputs(boxes_r, im_w=width, mode='box')
keypoints_flip = flip_inputs(keypoints_r, im_w=width)
# New right
keypoints_r_flip = flip_inputs(keypoints, im_w=width)
# combine the 2 modes
all_boxes_gt = [boxes_gt, boxes_gt_flip]
all_ys = [ys, ys_flip]
all_boxes = [boxes, boxes_flip]
all_keypoints = [keypoints, keypoints_flip]
all_keypoints_r = [keypoints_r, keypoints_r_flip]
else:
all_boxes_gt, all_ys = [boxes_gt], [ys]
all_boxes, all_keypoints = [boxes], [keypoints]
all_keypoints_r = [keypoints_r]
# Match each set of keypoint with a ground truth
self.dic_jo[phase]['K'].append(kk)
for ii, boxes_gt in enumerate(all_boxes_gt):
keypoints, keypoints_r = torch.tensor(all_keypoints[ii]), torch.tensor(all_keypoints_r[ii])
ys = all_ys[ii]
matches = get_iou_matches(all_boxes[ii], boxes_gt, self.iou_min)
for (idx, idx_gt) in matches:
keypoint = keypoints[idx:idx + 1]
lab = ys[idx_gt][:-1]
# Preprocess MonoLoco++
if self.mode == 'mono':
inp = preprocess_monoloco(keypoint, kk).view(-1).tolist()
lab = normalize_hwl(lab)
if ys[idx_gt][10] < 0.5:
self.dic_jo[phase]['kps'].append(keypoint.tolist())
self.dic_jo[phase]['X'].append(inp)
self.dic_jo[phase]['Y'].append(lab)
self.dic_jo[phase]['names'].append(name) # One image name for each annotation
append_cluster(self.dic_jo, phase, inp, lab, keypoint.tolist())
cnt_mono[phase] += 1
cnt_tot += 1
# Preprocess MonStereo
else:
zz = ys[idx_gt][2]
stereo_matches, cnt_amb = extract_stereo_matches(keypoint, keypoints_r, zz,
phase=phase, seed=cnt_pair_tot)
cnt_match_l += 1 if ii < 0.1 else 0 # matched instances
cnt_match_r += 1 if ii > 0.9 else 0
cnt_ambiguous += cnt_amb
# Monitor precision of classes
if phase == 'val':
if ys[idx_gt][10] == cat[idx] == 1:
correct_byc += 1
elif ys[idx_gt][10] == cat[idx] == 0:
correct_ped += 1
elif ys[idx_gt][10] != cat[idx] and ys[idx_gt][10] == 1:
wrong_byc += 1
elif ys[idx_gt][10] != cat[idx] and ys[idx_gt][10] == 0:
wrong_ped += 1
cnt_cyclist += 1 if ys[idx_gt][10] == 1 else 0
for num, (idx_r, s_match) in enumerate(stereo_matches):
label = ys[idx_gt][:-1] + [s_match]
if s_match > 0.9:
cnt_pair += 1
# Remove noise of very far instances for validation
# if (phase == 'val') and (ys[idx_gt][3] >= 50):
# continue
# ---> Save only positives unless there is no positive (keep positive flip and augm)
# if num > 0 and s_match < 0.9:
# continue
# Height augmentation
cnt_pair_tot += 1
cnt_extra_pair += 1 if ii == 1 else 0
flag_aug = False
if phase == 'train' and 3 < label[2] < 30 and s_match > 0.9:
flag_aug = True
elif phase == 'train' and 3 < label[2] < 30 and cnt_pair_tot % 2 == 0:
flag_aug = True
# Remove height augmentation
# flag_aug = False
if flag_aug:
kps_aug, labels_aug = height_augmentation(
keypoints[idx:idx+1], keypoints_r[idx_r:idx_r+1], label, s_match,
seed=cnt_pair_tot)
else:
kps_aug = [(keypoints[idx:idx+1], keypoints_r[idx_r:idx_r+1])]
labels_aug = [label]
for i, lab in enumerate(labels_aug):
(kps, kps_r) = kps_aug[i]
input_l = preprocess_monoloco(kps, kk).view(-1)
input_r = preprocess_monoloco(kps_r, kk).view(-1)
keypoint = torch.cat((kps, kps_r), dim=2).tolist()
inp = torch.cat((input_l, input_l - input_r)).tolist()
# Only relative distances
# inp_x = input[::2]
# inp = torch.cat((inp_x, input - input_r)).tolist()
# lab = normalize_hwl(lab)
if ys[idx_gt][10] < 0.5:
self.dic_jo[phase]['kps'].append(keypoint)
self.dic_jo[phase]['X'].append(inp)
self.dic_jo[phase]['Y'].append(lab)
self.dic_jo[phase]['names'].append(name) # One image name for each annotation
append_cluster(self.dic_jo, phase, inp, lab, keypoint)
cnt_tot += 1
if s_match > 0.9:
cnt_stereo[phase] += 1
else:
cnt_mono[phase] += 1
with open(self.path_joints, 'w') as file:
json.dump(self.dic_jo, file)
with open(os.path.join(self.path_names), 'w') as file:
json.dump(self.dic_names, file)
# cout
print(cnt_30)
print(cnt_less_30)
print('-' * 120)
print("Number of GT files: {}. Files with at least one pedestrian: {}. Files not found: {}"
.format(cnt_files, cnt_files_ped, cnt_fnf))
print("Ground truth matches : {:.1f} % for left images (train and val) and {:.1f} % for right images (train)"
.format(100*cnt_match_l / (cnt_gt['train'] + cnt_gt['val']), 100*cnt_match_r / cnt_gt['train']))
print("Total annotations: {}".format(cnt_tot))
print("Total number of cyclists: {}\n".format(cnt_cyclist))
print("Ambiguous instances removed: {}".format(cnt_ambiguous))
print("Extra pairs created with horizontal flipping: {}\n".format(cnt_extra_pair))
if self.mode == 'stereo':
print('Instances with stereo correspondence: {:.1f}% '.format(100 * cnt_pair / cnt_pair_tot))
for phase in ['train', 'val']:
cnt = cnt_mono[phase] + cnt_stereo[phase]
print("{}: annotations: {}. Stereo pairs {:.1f}% "
.format(phase.upper(), cnt, 100 * cnt_stereo[phase] / cnt))
print("\nOutput files:\n{}\n{}".format(self.path_names, self.path_joints))
print('-' * 120)
def prep_activity(self):
"""Augment ground-truth with flag activity"""
from monoloco.activity import social_interactions
main_dir = os.path.join('data', 'kitti')
dir_gt = os.path.join(main_dir, 'gt')
dir_out = os.path.join(main_dir, 'gt_activity')
make_new_directory(dir_out)
cnt_tp, cnt_tn = 0, 0
# Extract validation images for evaluation
category = 'pedestrian'
for name in self.set_val:
# Read
path_gt = os.path.join(dir_gt, name)
boxes_gt, ys, truncs_gt, occs_gt, lines = parse_ground_truth(path_gt, category, spherical=False,
verbose=True)
angles = [y[10] for y in ys]
dds = [y[4] for y in ys]
xz_centers = [[y[0], y[2]] for y in ys]
# Write
path_out = os.path.join(dir_out, name)
with open(path_out, "w+") as ff:
for idx, line in enumerate(lines):
if social_interactions(idx, xz_centers, angles, dds,
n_samples=1,
threshold_dist=self.THRESHOLD_DIST,
radii=self.RADII,
social_distance=self.SOCIAL_DISTANCE):
activity = '1'
cnt_tp += 1
else:
activity = '0'
cnt_tn += 1
line_new = line[:-1] + ' ' + activity + line[-1]
ff.write(line_new)
print(f'Written {len(self.set_val)} new files in {dir_out}')
print(f'Saved {cnt_tp} positive and {cnt_tn} negative annotations')
def _factory_phase(self, name):
"""Choose the phase"""
phase = None
flag = False
if name in self.set_train:
phase = 'train'
elif name in self.set_val:
phase = 'val'
else:
flag = True
return phase, flag
def crop_and_draw(im, box, keypoint):
box = [round(el) for el in box[:-1]]
center = (int((keypoint[0][0])), int((keypoint[1][0])))
radius = round((box[3]-box[1]) / 20)
im = cv2.circle(im, center, radius, color=(0, 255, 0), thickness=1)
crop = im[box[1]:box[3], box[0]:box[2]]
h_crop = crop.shape[0]
w_crop = crop.shape[1]
return crop, h_crop, w_crop

132
monoloco/prep/preprocess_ki.py
View File

@ -1,132 +0,0 @@
"""Preprocess annotations with KITTI ground-truth"""
import os
import glob
import copy
import logging
from collections import defaultdict
import json
import datetime
from .transforms import transform_keypoints
from ..utils import get_calibration, split_training, parse_ground_truth, get_iou_matches, append_cluster
from ..network.process import preprocess_pifpaf, preprocess_monoloco
class PreprocessKitti:
"""Prepare arrays with same format as nuScenes preprocessing but using ground truth txt files"""
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
dic_jo = {'train': dict(X=[], Y=[], names=[], kps=[], boxes_3d=[], K=[],
clst=defaultdict(lambda: defaultdict(list))),
'val': dict(X=[], Y=[], names=[], kps=[], boxes_3d=[], K=[],
clst=defaultdict(lambda: defaultdict(list))),
'test': dict(X=[], Y=[], names=[], kps=[], boxes_3d=[], K=[],
clst=defaultdict(lambda: defaultdict(list)))}
dic_names = defaultdict(lambda: defaultdict(list))
def __init__(self, dir_ann, iou_min):
self.dir_ann = dir_ann
self.iou_min = iou_min
self.dir_gt = os.path.join('data', 'kitti', 'gt')
self.names_gt = tuple(os.listdir(self.dir_gt))
self.dir_kk = os.path.join('data', 'kitti', 'calib')
self.list_gt = glob.glob(self.dir_gt + '/*.txt')
assert os.path.exists(self.dir_gt), "Ground truth dir does not exist"
assert os.path.exists(self.dir_ann), "Annotation dir does not exist"
now = datetime.datetime.now()
now_time = now.strftime("%Y%m%d-%H%M")[2:]
dir_out = os.path.join('data', 'arrays')
self.path_joints = os.path.join(dir_out, 'joints-kitti-' + now_time + '.json')
self.path_names = os.path.join(dir_out, 'names-kitti-' + now_time + '.json')
path_train = os.path.join('splits', 'kitti_train.txt')
path_val = os.path.join('splits', 'kitti_val.txt')
self.set_train, self.set_val = split_training(self.names_gt, path_train, path_val)
def run(self):
"""Save json files"""
cnt_gt = cnt_files = cnt_files_ped = cnt_fnf = 0
dic_cnt = {'train': 0, 'val': 0, 'test': 0}
for name in self.names_gt:
path_gt = os.path.join(self.dir_gt, name)
basename, _ = os.path.splitext(name)
phase, flag = self._factory_phase(name)
if flag:
cnt_fnf += 1
continue
# Extract keypoints
path_txt = os.path.join(self.dir_kk, basename + '.txt')
p_left, _ = get_calibration(path_txt)
kk = p_left[0]
# Iterate over each line of the gt file and save box location and distances
boxes_gt, boxes_3d, dds_gt = parse_ground_truth(path_gt, category='all')[:3]
self.dic_names[basename + '.png']['boxes'] = copy.deepcopy(boxes_gt)
self.dic_names[basename + '.png']['dds'] = copy.deepcopy(dds_gt)
self.dic_names[basename + '.png']['K'] = copy.deepcopy(kk)
cnt_gt += len(boxes_gt)
cnt_files += 1
cnt_files_ped += min(len(boxes_gt), 1) # if no boxes 0 else 1
# Find the annotations if exists
try:
with open(os.path.join(self.dir_ann, basename + '.png.pifpaf.json'), 'r') as f:
annotations = json.load(f)
boxes, keypoints = preprocess_pifpaf(annotations, im_size=(1238, 374))
keypoints_hflip = transform_keypoints(keypoints, mode='flip')
inputs = preprocess_monoloco(keypoints, kk).tolist()
inputs_hflip = preprocess_monoloco(keypoints_hflip, kk).tolist()
all_keypoints = [keypoints, keypoints_hflip]
all_inputs = [inputs, inputs_hflip]
except FileNotFoundError:
boxes = []
# Match each set of keypoint with a ground truth
matches = get_iou_matches(boxes, boxes_gt, self.iou_min)
for (idx, idx_gt) in matches:
for nn, keypoints in enumerate(all_keypoints):
inputs = all_inputs[nn]
self.dic_jo[phase]['kps'].append(keypoints[idx])
self.dic_jo[phase]['X'].append(inputs[idx])
self.dic_jo[phase]['Y'].append([dds_gt[idx_gt]]) # Trick to make it (nn,1)
self.dic_jo[phase]['boxes_3d'].append(boxes_3d[idx_gt])
self.dic_jo[phase]['K'].append(kk)
self.dic_jo[phase]['names'].append(name) # One image name for each annotation
append_cluster(self.dic_jo, phase, inputs[idx], dds_gt[idx_gt], keypoints[idx])
dic_cnt[phase] += 1
with open(self.path_joints, 'w') as file:
json.dump(self.dic_jo, file)
with open(os.path.join(self.path_names), 'w') as file:
json.dump(self.dic_names, file)
for phase in ['train', 'val', 'test']:
print("Saved {} annotations for phase {}"
.format(dic_cnt[phase], phase))
print("Number of GT files: {}. Files with at least one pedestrian: {}. Files not found: {}"
.format(cnt_files, cnt_files_ped, cnt_fnf))
print("Matched : {:.1f} % of the ground truth instances"
.format(100 * (dic_cnt['train'] + dic_cnt['val']) / cnt_gt))
print("\nOutput files:\n{}\n{}\n".format(self.path_names, self.path_joints))
def _factory_phase(self, name):
"""Choose the phase"""
phase = None
flag = False
if name in self.set_train:
phase = 'train'
elif name in self.set_val:
phase = 'val'
else:
flag = True
return phase, flag

392
monoloco/prep/preprocess_kitti.py Normal file
View File

@ -0,0 +1,392 @@
# pylint: disable=too-many-statements, too-many-branches, too-many-nested-blocks
"""Preprocess annotations with KITTI ground-truth"""
import os
import glob
import copy
import math
import logging
from collections import defaultdict
import json
import warnings
import datetime
from PIL import Image
import torch
from .. import __version__
from ..utils import split_training, get_iou_matches, append_cluster, get_calibration, open_annotations, \
extract_stereo_matches, make_new_directory, \
check_conditions, to_spherical, correct_angle
from ..network.process import preprocess_pifpaf, preprocess_monoloco
from .transforms import flip_inputs, flip_labels, height_augmentation
class PreprocessKitti:
"""Prepare arrays with same format as nuScenes preprocessing but using ground truth txt files"""
# KITTI Dataset files
dir_gt = os.path.join('data', 'kitti', 'gt')
dir_images = os.path.join('data', 'kitti', 'images')
dir_kk = os.path.join('data', 'kitti', 'calib')
# SOCIAL DISTANCING PARAMETERS
THRESHOLD_DIST = 2 # Threshold to check distance of people
RADII = (0.3, 0.5, 1) # expected radii of the o-space
SOCIAL_DISTANCE = True
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
dic_jo = {
'train': dict(X=[], Y=[], names=[], kps=[], K=[], clst=defaultdict(lambda: defaultdict(list))),
'val': dict(X=[], Y=[], names=[], kps=[], K=[], clst=defaultdict(lambda: defaultdict(list))),
'test': dict(X=[], Y=[], names=[], kps=[], K=[], clst=defaultdict(lambda: defaultdict(list))),
'version': __version__,
}
dic_names = defaultdict(lambda: defaultdict(list))
dic_std = defaultdict(lambda: defaultdict(list))
categories_gt = dict(train=['Pedestrian', 'Person_sitting'], val=['Pedestrian'])
def __init__(self, dir_ann, mode='mono', iou_min=0.3, sample=False):
self.dir_ann = dir_ann
self.mode = mode
self.iou_min = iou_min
self.sample = sample
assert os.path.isdir(self.dir_ann), "Annotation directory not found"
assert any(os.scandir(self.dir_ann)), "Annotation directory empty"
assert os.path.isdir(self.dir_gt), "Ground truth directory not found"
assert any(os.scandir(self.dir_gt)), "Ground-truth directory empty"
if self.mode == 'stereo':
assert os.path.isdir(self.dir_ann + '_right'), "Annotation directory for right images not found"
assert any(os.scandir(self.dir_ann + '_right')), "Annotation directory for right images empty"
elif not os.path.isdir(self.dir_ann + '_right') or not any(os.scandir(self.dir_ann + '_right')):
warnings.warn('Horizontal flipping not applied as annotation directory for right images not found/empty')
assert self.mode in ('mono', 'stereo'), "modality not recognized"
self.names_gt = tuple(os.listdir(self.dir_gt))
self.list_gt = glob.glob(self.dir_gt + '/*.txt')
now = datetime.datetime.now()
now_time = now.strftime("%Y%m%d-%H%M")[2:]
dir_out = os.path.join('data', 'arrays')
self.path_joints = os.path.join(dir_out, 'joints-kitti-' + self.mode + '-' + now_time + '.json')
self.path_names = os.path.join(dir_out, 'names-kitti-' + self.mode + '-' + now_time + '.json')
path_train = os.path.join('splits', 'kitti_train.txt')
path_val = os.path.join('splits', 'kitti_val.txt')
self.set_train, self.set_val = split_training(self.names_gt, path_train, path_val)
self.phase, self.name = None, None
self.stats = defaultdict(int)
self.stats_stereo = defaultdict(int)
def run(self):
# self.names_gt = ('002282.txt',)
for self.name in self.names_gt:
# Extract ground truth
path_gt = os.path.join(self.dir_gt, self.name)
basename, _ = os.path.splitext(self.name)
self.phase, file_not_found = self._factory_phase(self.name)
category = 'all' if self.phase == 'train' else 'pedestrian'
if file_not_found:
self.stats['fnf'] += 1
continue
boxes_gt, labels, _, _, _ = parse_ground_truth(path_gt, category=category, spherical=True)
self.stats['gt_' + self.phase] += len(boxes_gt)
self.stats['gt_files'] += 1
self.stats['gt_files_ped'] += min(len(boxes_gt), 1) # if no boxes 0 else 1
self.dic_names[basename + '.png']['boxes'] = copy.deepcopy(boxes_gt)
self.dic_names[basename + '.png']['ys'] = copy.deepcopy(labels)
# Extract annotations
dic_boxes, dic_kps, dic_gt = self.parse_annotations(boxes_gt, labels, basename)
if dic_boxes is None: # No annotations
continue
self.dic_names[basename + '.png']['K'] = copy.deepcopy(dic_gt['K'])
self.dic_jo[self.phase]['K'].append(dic_gt['K'])
# Match each set of keypoint with a ground truth
for ii, boxes_gt in enumerate(dic_boxes['gt']):
kps, kps_r = torch.tensor(dic_kps['left'][ii]), torch.tensor(dic_kps['right'][ii])
matches = get_iou_matches(dic_boxes['left'][ii], boxes_gt, self.iou_min)
self.stats['flipping_match'] += len(matches) if ii == 1 else 0
for (idx, idx_gt) in matches:
cat_gt = dic_gt['labels'][ii][idx_gt][-1]
if cat_gt not in self.categories_gt[self.phase]: # only for training as cyclists are also extracted
continue
kp = kps[idx:idx + 1]
kk = dic_gt['K']
label = dic_gt['labels'][ii][idx_gt][:-1]
self.stats['match'] += 1
assert len(label) == 10, 'dimensions of monocular label is wrong'
if self.mode == 'mono':
self._process_annotation_mono(kp, kk, label)
else:
self._process_annotation_stereo(kp, kk, label, kps_r)
with open(self.path_joints, 'w') as file:
json.dump(self.dic_jo, file)
with open(os.path.join(self.path_names), 'w') as file:
json.dump(self.dic_names, file)
self._cout()
def parse_annotations(self, boxes_gt, labels, basename):
path_im = os.path.join(self.dir_images, basename + '.png')
path_calib = os.path.join(self.dir_kk, basename + '.txt')
min_conf = 0 if self.phase == 'train' else 0.1
# Check image size
with Image.open(path_im) as im:
width, height = im.size
# Extract left keypoints
annotations, kk, _ = factory_file(path_calib, self.dir_ann, basename)
boxes, keypoints = preprocess_pifpaf(annotations, im_size=(width, height), min_conf=min_conf)
if not keypoints:
return None, None, None
# Stereo-based horizontal flipping for training (obtaining ground truth for right images)
self.stats['instances'] += len(keypoints)
annotations_r, _, _ = factory_file(path_calib, self.dir_ann, basename, ann_type='right')
boxes_r, keypoints_r = preprocess_pifpaf(annotations_r, im_size=(width, height), min_conf=min_conf)
if not keypoints_r: # Duplicate the left one(s)
all_boxes_gt, all_labels = [boxes_gt], [labels]
boxes_r, keypoints_r = boxes[0:1].copy(), keypoints[0:1].copy()
all_boxes, all_keypoints = [boxes], [keypoints]
all_keypoints_r = [keypoints_r]
elif self.phase == 'train':
# GT)
boxes_gt_flip, ys_flip = flip_labels(boxes_gt, labels, im_w=width)
# New left
boxes_flip = flip_inputs(boxes_r, im_w=width, mode='box')
keypoints_flip = flip_inputs(keypoints_r, im_w=width)
# New right
keypoints_r_flip = flip_inputs(keypoints, im_w=width)
# combine the 2 modes
all_boxes_gt = [boxes_gt, boxes_gt_flip]
all_labels = [labels, ys_flip]
all_boxes = [boxes, boxes_flip]
all_keypoints = [keypoints, keypoints_flip]
all_keypoints_r = [keypoints_r, keypoints_r_flip]
else:
all_boxes_gt, all_labels = [boxes_gt], [labels]
all_boxes, all_keypoints = [boxes], [keypoints]
all_keypoints_r = [keypoints_r]
dic_boxes = dict(left=all_boxes, gt=all_boxes_gt)
dic_kps = dict(left=all_keypoints, right=all_keypoints_r)
dic_gt = dict(K=kk, labels=all_labels)
return dic_boxes, dic_kps, dic_gt
def _process_annotation_mono(self, kp, kk, label):
"""For a single annotation, process all the labels and save them"""
kp = kp.tolist()
inp = preprocess_monoloco(kp, kk).view(-1).tolist()
# Save
self.dic_jo[self.phase]['kps'].append(kp)
self.dic_jo[self.phase]['X'].append(inp)
self.dic_jo[self.phase]['Y'].append(label)
self.dic_jo[self.phase]['names'].append(self.name) # One image name for each annotation
append_cluster(self.dic_jo, self.phase, inp, label, kp)
self.stats['total_' + self.phase] += 1
def _process_annotation_stereo(self, kp, kk, label, kps_r):
"""For a reference annotation, combine it with some (right) annotations and save it"""
zz = label[2]
stereo_matches, cnt_amb = extract_stereo_matches(kp, kps_r, zz,
phase=self.phase,
seed=self.stats_stereo['pair'])
self.stats_stereo['ambiguous'] += cnt_amb
for idx_r, s_match in stereo_matches:
label_s = label + [s_match] # add flag to distinguish "true pairs and false pairs"
self.stats_stereo['true_pair'] += 1 if s_match > 0.9 else 0
self.stats_stereo['pair'] += 1 # before augmentation
# ---> Remove noise of very far instances for validation
# if (self.phase == 'val') and (label[3] >= 50):
# continue
# ---> Save only positives unless there is no positive (keep positive flip and augm)
# if num > 0 and s_match < 0.9:
# continue
# Height augmentation
flag_aug = False
if self.phase == 'train' and 3 < label[2] < 30 and (s_match > 0.9 or self.stats_stereo['pair'] % 2 == 0):
flag_aug = True
# Remove height augmentation
# flag_aug = False
if flag_aug:
kps_aug, labels_aug = height_augmentation(kp, kps_r[idx_r:idx_r + 1], label_s,
seed=self.stats_stereo['pair'])
else:
kps_aug = [(kp, kps_r[idx_r:idx_r + 1])]
labels_aug = [label_s]
for i, lab in enumerate(labels_aug):
assert len(lab) == 11, 'dimensions of stereo label is wrong'
self.stats_stereo['pair_aug'] += 1
(kp_aug, kp_aug_r) = kps_aug[i]
input_l = preprocess_monoloco(kp_aug, kk).view(-1)
input_r = preprocess_monoloco(kp_aug_r, kk).view(-1)
keypoint = torch.cat((kp_aug, kp_aug_r), dim=2).tolist()
inp = torch.cat((input_l, input_l - input_r)).tolist()
self.dic_jo[self.phase]['kps'].append(keypoint)
self.dic_jo[self.phase]['X'].append(inp)
self.dic_jo[self.phase]['Y'].append(lab)
self.dic_jo[self.phase]['names'].append(self.name) # One image name for each annotation
append_cluster(self.dic_jo, self.phase, inp, lab, keypoint)
self.stats_stereo['total_' + self.phase] += 1 # including height augmentation
def _cout(self):
print('-' * 100)
print(f"Number of GT files: {self.stats['gt_files']} ")
print(f"Files with at least one pedestrian/cyclist: {self.stats['gt_files_ped']}")
print(f"Files not found: {self.stats['fnf']}")
print('-' * 100)
our = self.stats['match'] - self.stats['flipping_match']
gt = self.stats['gt_train'] + self.stats['gt_val']
print(f"Ground truth matches: {100 * our / gt:.1f} for left images (train and val)")
print(f"Parsed instances: {self.stats['instances']}")
print(f"Ground truth instances: {gt}")
print(f"Matched instances: {our}")
print(f"Including horizontal flipping: {self.stats['match']}")
if self.mode == 'stereo':
print('-' * 100)
print(f"Ambiguous instances removed: {self.stats_stereo['ambiguous']}")
print(f"True pairs ratio: {100 * self.stats_stereo['true_pair'] / self.stats_stereo['pair']:.1f}% ")
print(f"Height augmentation pairs: {self.stats_stereo['pair_aug'] - self.stats_stereo['pair']} ")
print('-' * 100)
total_train = self.stats_stereo['total_train'] if self.mode == 'stereo' else self.stats['total_train']
total_val = self.stats_stereo['total_val'] if self.mode == 'stereo' else self.stats['total_val']
print(f"Total annotations for TRAINING: {total_train}")
print(f"Total annotations for VALIDATION: {total_val}")
print('-' * 100)
print(f"\nOutput files:\n{self.path_names}\n{self.path_joints}")
print('-' * 100)
def process_activity(self):
"""Augment ground-truth with flag activity"""
from monoloco.activity import social_interactions # pylint: disable=import-outside-toplevel
main_dir = os.path.join('data', 'kitti')
dir_gt = os.path.join(main_dir, 'gt')
dir_out = os.path.join(main_dir, 'gt_activity')
make_new_directory(dir_out)
cnt_tp, cnt_tn = 0, 0
# Extract validation images for evaluation
category = 'pedestrian'
for name in self.set_val:
# Read
path_gt = os.path.join(dir_gt, name)
_, ys, _, _, lines = parse_ground_truth(path_gt, category, spherical=False)
angles = [y[10] for y in ys]
dds = [y[4] for y in ys]
xz_centers = [[y[0], y[2]] for y in ys]
# Write
path_out = os.path.join(dir_out, name)
with open(path_out, "w+") as ff:
for idx, line in enumerate(lines):
if social_interactions(idx, xz_centers, angles, dds,
n_samples=1,
threshold_dist=self.THRESHOLD_DIST,
radii=self.RADII,
social_distance=self.SOCIAL_DISTANCE):
activity = '1'
cnt_tp += 1
else:
activity = '0'
cnt_tn += 1
line_new = line[:-1] + ' ' + activity + line[-1]
ff.write(line_new)
print(f'Written {len(self.set_val)} new files in {dir_out}')
print(f'Saved {cnt_tp} positive and {cnt_tn} negative annotations')
def _factory_phase(self, name):
"""Choose the phase"""
phase = None
flag = False
if name in self.set_train:
phase = 'train'
elif name in self.set_val:
phase = 'val'
else:
flag = True
return phase, flag
def parse_ground_truth(path_gt, category, spherical=False):
"""Parse KITTI ground truth files"""
boxes_gt = []
labels = []
truncs_gt = [] # Float from 0 to 1
occs_gt = [] # Either 0,1,2,3 fully visible, partly occluded, largely occluded, unknown
lines = []
with open(path_gt, "r") as f_gt:
for line_gt in f_gt:
line = line_gt.split()
if not check_conditions(line_gt, category, method='gt'):
continue
truncs_gt.append(float(line[1]))
occs_gt.append(int(line[2]))
boxes_gt.append([float(x) for x in line[4:8]])
xyz = [float(x) for x in line[11:14]]
hwl = [float(x) for x in line[8:11]]
dd = float(math.sqrt(xyz[0] ** 2 + xyz[1] ** 2 + xyz[2] ** 2))
yaw = float(line[14])
assert - math.pi <= yaw <= math.pi
alpha = float(line[3])
sin, cos, yaw_corr = correct_angle(yaw, xyz)
assert min(abs(-yaw_corr - alpha), (abs(yaw_corr - alpha))) < 0.15, "more than 10 degrees of error"
if spherical:
rtp = to_spherical(xyz)
loc = rtp[1:3] + xyz[2:3] + rtp[0:1] # [theta, psi, z, r]
else:
loc = xyz + [dd]
cat = line[0] # 'Pedestrian', or 'Person_sitting' for people
output = loc + hwl + [sin, cos, yaw, cat]
labels.append(output)
lines.append(line_gt)
return boxes_gt, labels, truncs_gt, occs_gt, lines
def factory_file(path_calib, dir_ann, basename, ann_type='left'):
"""Choose the annotation and the calibration files"""
assert ann_type in ('left', 'right')
p_left, p_right = get_calibration(path_calib)
if ann_type == 'left':
kk, tt = p_left[:]
path_ann = os.path.join(dir_ann, basename + '.png.predictions.json')
# The right folder is called <NameOfLeftFolder>_right
else:
kk, tt = p_right[:]
path_ann = os.path.join(dir_ann + '_right', basename + '.png.predictions.json')
annotations = open_annotations(path_ann)
return annotations, kk, tt

17
monoloco/prep/preprocess_nu.py
View File

@ -150,7 +150,6 @@ def extract_ground_truth(boxes_obj, kk, spherical=True):
ys = [] ys = []
for box_obj in boxes_obj: for box_obj in boxes_obj:
# Select category # Select category
if box_obj.name[:6] != 'animal': if box_obj.name[:6] != 'animal':
general_name = box_obj.name.split('.')[0] + '.' + box_obj.name.split('.')[1] general_name = box_obj.name.split('.')[0] + '.' + box_obj.name.split('.')[1]
@ -248,8 +247,6 @@ def extract_social(inputs, ys, keypoints, idx, matches):
all_inputs.extend(inputs[idx]) all_inputs.extend(inputs[idx])
indices_idx = [idx for (idx, idx_gt) in matches] indices_idx = [idx for (idx, idx_gt) in matches]
if len(sorted_indices) > 2:
aa = 5
for ii in range(1, 3): for ii in range(1, 3):
try: try:
index = sorted_indices[ii] index = sorted_indices[ii]
@ -266,17 +263,3 @@ def extract_social(inputs, ys, keypoints, idx, matches):
all_inputs.extend([0.] * 2) all_inputs.extend([0.] * 2)
assert len(all_inputs) == 34 + 2 * 2 assert len(all_inputs) == 34 + 2 * 2
return all_inputs return all_inputs
# def get_jean_yaw(box_obj):
# b_corners = box_obj.bottom_corners()
# center = box_obj.center
# back_point = [(b_corners[0, 2] + b_corners[0, 3]) / 2, (b_corners[2, 2] + b_corners[2, 3]) / 2]
#
# x = b_corners[0, :] - back_point[0]
# y = b_corners[2, :] - back_point[1]
#
# angle = math.atan2((x[0] + x[1]) / 2, (y[0] + y[1]) / 2) * 180 / 3.14
# angle = (angle + 360) % 360
# correction = math.atan2(center[0], center[2]) * 180 / 3.14
# return angle, correction

21
monoloco/prep/transforms.py
View File

@ -1,8 +1,11 @@
import math import math
from copy import deepcopy from copy import deepcopy
import numpy as np import numpy as np
from ..utils import correct_angle, to_cartesian, to_spherical
BASELINE = 0.54 BASELINE = 0.54
BF = BASELINE * 721 BF = BASELINE * 721
@ -75,7 +78,6 @@ def flip_inputs(keypoints, im_w, mode=None):
def flip_labels(boxes_gt, labels, im_w): def flip_labels(boxes_gt, labels, im_w):
"""Correct x, d positions and angles after horizontal flipping""" """Correct x, d positions and angles after horizontal flipping"""
from ..utils import correct_angle, to_cartesian, to_spherical
boxes_flip = deepcopy(boxes_gt) boxes_flip = deepcopy(boxes_gt)
labels_flip = deepcopy(labels) labels_flip = deepcopy(labels)
@ -98,29 +100,28 @@ def flip_labels(boxes_gt, labels, im_w):
yaw = label_flip[9] yaw = label_flip[9]
yaw_n = math.copysign(1, yaw) * (np.pi - abs(yaw)) # Horizontal flipping change of angle yaw_n = math.copysign(1, yaw) * (np.pi - abs(yaw)) # Horizontal flipping change of angle
sin, cos, yaw_corr = correct_angle(yaw_n, xyz) sin, cos, _ = correct_angle(yaw_n, xyz)
label_flip[7], label_flip[8], label_flip[9] = sin, cos, yaw_n label_flip[7], label_flip[8], label_flip[9] = sin, cos, yaw_n
return boxes_flip, labels_flip return boxes_flip, labels_flip
def height_augmentation(kps, kps_r, label, s_match, seed=0): def height_augmentation(kps, kps_r, label_s, seed=0):
""" """
label: theta, psi, z, rho, wlh, sin, cos, yaw, cat label_s: theta, psi, z, rho, wlh, sin, cos, s_match
""" """
from ..utils import to_cartesian n_labels = 3 if label_s[-1] > 0.9 else 1
n_labels = 3 if s_match > 0.9 else 1
height_min = 1.2 height_min = 1.2
height_max = 2 height_max = 2
av_height = 1.71 av_height = 1.71
kps_aug = [[kps.clone(), kps_r.clone()] for _ in range(n_labels+1)] kps_aug = [[kps.clone(), kps_r.clone()] for _ in range(n_labels+1)]
labels_aug = [label.copy() for _ in range(n_labels+1)] # Maintain the original labels_aug = [label_s.copy() for _ in range(n_labels+1)] # Maintain the original
np.random.seed(seed) np.random.seed(seed)
heights = np.random.uniform(height_min, height_max, n_labels) # 3 samples heights = np.random.uniform(height_min, height_max, n_labels) # 3 samples
zzs = heights * label[2] / av_height zzs = heights * label_s[2] / av_height
disp = BF / label[2] disp = BF / label_s[2]
rtp = label[3:4] + label[0:2] # Originally t,p,z,r rtp = label_s[3:4] + label_s[0:2] # Originally t,p,z,r
xyz = to_cartesian(rtp) xyz = to_cartesian(rtp)
for i in range(n_labels): for i in range(n_labels):

21
monoloco/run.py
View File

@ -1,4 +1,4 @@
# pylint: disable=too-many-branches, too-many-statements # pylint: disable=too-many-branches, too-many-statements, import-outside-toplevel
import argparse import argparse
@ -18,6 +18,7 @@ def cli():
# Predict (2D pose and/or 3D location from images) # Predict (2D pose and/or 3D location from images)
predict_parser.add_argument('images', nargs='*', help='input images') predict_parser.add_argument('images', nargs='*', help='input images')
predict_parser.add_argument('--glob', help='glob expression for input images (for many images)') predict_parser.add_argument('--glob', help='glob expression for input images (for many images)')
predict_parser.add_argument('--checkpoint', help='pifpaf model')
predict_parser.add_argument('-o', '--output-directory', help='Output directory') predict_parser.add_argument('-o', '--output-directory', help='Output directory')
predict_parser.add_argument('--output_types', nargs='+', default=['json'], predict_parser.add_argument('--output_types', nargs='+', default=['json'],
help='what to output: json keypoints skeleton for Pifpaf' help='what to output: json keypoints skeleton for Pifpaf'
@ -35,9 +36,10 @@ def cli():
predict_parser.add_argument('--instance-threshold', type=float, default=None, help='threshold for entire instance') predict_parser.add_argument('--instance-threshold', type=float, default=None, help='threshold for entire instance')
predict_parser.add_argument('--seed-threshold', type=float, default=0.5, help='threshold for single seed') predict_parser.add_argument('--seed-threshold', type=float, default=0.5, help='threshold for single seed')
predict_parser.add_argument('--disable-cuda', action='store_true', help='disable CUDA') predict_parser.add_argument('--disable-cuda', action='store_true', help='disable CUDA')
predict_parser.add_argument('--focal', predict_parser.add_argument('--precise-rescaling', dest='fast_rescaling', default=True, action='store_false',
help='focal length in mm for a sensor size of 7.2x5.4 mm. Default nuScenes sensor', help='use more exact image rescaling (requires scipy)')
type=float, default=5.7) predict_parser.add_argument('--decoder-workers', default=None, type=int,
help='number of workers for pose decoding, 0 for windows')
decoder.cli(parser) decoder.cli(parser)
logger.cli(parser) logger.cli(parser)
@ -53,6 +55,8 @@ def cli():
predict_parser.add_argument('--n_dropout', type=int, help='Epistemic uncertainty evaluation', default=0) predict_parser.add_argument('--n_dropout', type=int, help='Epistemic uncertainty evaluation', default=0)
predict_parser.add_argument('--dropout', type=float, help='dropout parameter', default=0.2) predict_parser.add_argument('--dropout', type=float, help='dropout parameter', default=0.2)
predict_parser.add_argument('--show_all', help='only predict ground-truth matches or all', action='store_true') predict_parser.add_argument('--show_all', help='only predict ground-truth matches or all', action='store_true')
predict_parser.add_argument('--focal', help='focal length in mm for a sensor size of 7.2x5.4 mm. (nuScenes)',
type=float, default=5.7)
# Social distancing and social interactions # Social distancing and social interactions
predict_parser.add_argument('--social_distance', help='social', action='store_true') predict_parser.add_argument('--social_distance', help='social', action='store_true')
@ -74,11 +78,12 @@ def cli():
# Training # Training
training_parser.add_argument('--joints', help='Json file with input joints', required=True) training_parser.add_argument('--joints', help='Json file with input joints', required=True)
training_parser.add_argument('--mode', help='mono, stereo', default='mono') training_parser.add_argument('--mode', help='mono, stereo', default='mono')
training_parser.add_argument('--out', help='output_path, e.g., data/outputs/test.pkl')
training_parser.add_argument('-e', '--epochs', type=int, help='number of epochs to train for', default=500) training_parser.add_argument('-e', '--epochs', type=int, help='number of epochs to train for', default=500)
training_parser.add_argument('--bs', type=int, default=512, help='input batch size') training_parser.add_argument('--bs', type=int, default=512, help='input batch size')
training_parser.add_argument('--monocular', help='whether to train monoloco', action='store_true') training_parser.add_argument('--monocular', help='whether to train monoloco', action='store_true')
training_parser.add_argument('--dropout', type=float, help='dropout. Default no dropout', default=0.2) training_parser.add_argument('--dropout', type=float, help='dropout. Default no dropout', default=0.2)
training_parser.add_argument('--lr', type=float, help='learning rate', default=0.001) training_parser.add_argument('--lr', type=float, help='learning rate', default=0.002)
training_parser.add_argument('--sched_step', type=float, help='scheduler step time (epochs)', default=30) training_parser.add_argument('--sched_step', type=float, help='scheduler step time (epochs)', default=30)
training_parser.add_argument('--sched_gamma', type=float, help='Scheduler multiplication every step', default=0.98) training_parser.add_argument('--sched_gamma', type=float, help='Scheduler multiplication every step', default=0.98)
training_parser.add_argument('--hidden_size', type=int, help='Number of hidden units in the model', default=1024) training_parser.add_argument('--hidden_size', type=int, help='Number of hidden units in the model', default=1024)
@ -131,10 +136,10 @@ def main():
prep = PreprocessNuscenes(args.dir_ann, args.dir_nuscenes, args.dataset, args.iou_min) prep = PreprocessNuscenes(args.dir_ann, args.dir_nuscenes, args.dataset, args.iou_min)
prep.run() prep.run()
else: else:
from .prep.prep_kitti import PreprocessKitti from .prep.preprocess_kitti import PreprocessKitti
prep = PreprocessKitti(args.dir_ann, mode=args.mode, iou_min=args.iou_min) prep = PreprocessKitti(args.dir_ann, mode=args.mode, iou_min=args.iou_min)
if args.activity: if args.activity:
prep.prep_activity() prep.process_activity()
else: else:
prep.run() prep.run()
@ -162,7 +167,7 @@ def main():
elif args.geometric: elif args.geometric:
assert args.joints, "joints argument not provided" assert args.joints, "joints argument not provided"
from .network.geom_baseline import geometric_baseline from .eval.geom_baseline import geometric_baseline
geometric_baseline(args.joints) geometric_baseline(args.joints)
elif args.variance: elif args.variance:

4
monoloco/train/datasets.py
View File

@ -60,6 +60,7 @@ class KeypointsDataset(Dataset):
self.outputs_all = torch.tensor(dic_jo[phase]['Y']) self.outputs_all = torch.tensor(dic_jo[phase]['Y'])
self.names_all = dic_jo[phase]['names'] self.names_all = dic_jo[phase]['names']
self.kps_all = torch.tensor(dic_jo[phase]['kps']) self.kps_all = torch.tensor(dic_jo[phase]['kps'])
self.version = dic_jo['version']
# Extract annotations divided in clusters # Extract annotations divided in clusters
self.dic_clst = dic_jo[phase]['clst'] self.dic_clst = dic_jo[phase]['clst']
@ -90,3 +91,6 @@ class KeypointsDataset(Dataset):
count = len(self.dic_clst[clst]['Y']) count = len(self.dic_clst[clst]['Y'])
return inputs, outputs, count return inputs, outputs, count
def get_version(self):
return self.version

60
monoloco/train/losses.py
View File

@ -1,9 +1,15 @@
"""Inspired by Openpifpaf""" """
Adapted from https://github.com/openpifpaf,
which is: 'Copyright 2019-2021 by Sven Kreiss and contributors. All rights reserved.'
and licensed under GNU AGPLv3
"""
import math import math
import torch import torch
import torch.nn as nn import torch.nn as nn
import numpy as np import numpy as np
import matplotlib.pyplot as plt
from ..network import extract_labels, extract_labels_aux, extract_outputs from ..network import extract_labels, extract_labels_aux, extract_outputs
@ -180,6 +186,58 @@ class GaussianLoss(torch.nn.Module):
return torch.sum(values) return torch.sum(values)
class CustomL1Loss(torch.nn.Module):
"""
Experimental, not used.
L1 loss with more weight to errors at a shorter distance
It inherits from nn.module so it supports backward
"""
def __init__(self, dic_norm, device, beta=1):
super().__init__()
self.dic_norm = dic_norm
self.device = device
self.beta = beta
@staticmethod
def compute_weights(xx, beta=1):
"""
Return the appropriate weight depending on the distance and the hyperparameter chosen
alpha = 1 refers to the curve of A Photogrammetric Approach for Real-time...
It is made for unnormalized outputs (to be more understandable)
From 70 meters on every value is weighted the same (0.1**beta)
Alpha is optional value from Focal loss. Yet to be analyzed
"""
# alpha = np.maximum(1, 10 ** (beta - 1))
alpha = 1
ww = np.maximum(0.1, 1 - xx / 78)**beta
return alpha * ww
def print_loss(self):
xx = np.linspace(0, 80, 100)
y1 = self.compute_weights(xx, beta=1)
y2 = self.compute_weights(xx, beta=2)
y3 = self.compute_weights(xx, beta=3)
plt.plot(xx, y1)
plt.plot(xx, y2)
plt.plot(xx, y3)
plt.xlabel("Distance [m]")
plt.ylabel("Loss function Weight")
plt.legend(("Beta = 1", "Beta = 2", "Beta = 3"))
plt.show()
def forward(self, output, target):
unnormalized_output = output.cpu().detach().numpy() * self.dic_norm['std']['Y'] + self.dic_norm['mean']['Y']
weights_np = self.compute_weights(unnormalized_output, self.beta)
weights = torch.from_numpy(weights_np).float().to(self.device) # To make weights in the same cuda device
losses = torch.abs(output - target) * weights
loss = losses.mean() # Mean over the batch
return loss
def angle_loss(orient, gt_orient): def angle_loss(orient, gt_orient):
"""Only for evaluation""" """Only for evaluation"""
angles = torch.atan2(orient[:, 0], orient[:, 1]) angles = torch.atan2(orient[:, 0], orient[:, 1])

125
monoloco/train/trainer.py
View File

@ -1,8 +1,10 @@
# pylint: disable=too-many-statements # pylint: disable=too-many-statements
""" """
Training and evaluation of a neural network which predicts 3D localization and confidence intervals Training and evaluation of a neural network that, given 2D joints, estimates:
given 2d joints - 3D localization and confidence intervals
- Orientation
- Bounding box dimensions
""" """
import copy import copy
@ -12,7 +14,6 @@ import logging
from collections import defaultdict from collections import defaultdict
import sys import sys
import time import time
import warnings
from itertools import chain from itertools import chain
import matplotlib.pyplot as plt import matplotlib.pyplot as plt
@ -20,10 +21,11 @@ import torch
from torch.utils.data import DataLoader from torch.utils.data import DataLoader
from torch.optim import lr_scheduler from torch.optim import lr_scheduler
from .. import __version__
from .datasets import KeypointsDataset from .datasets import KeypointsDataset
from .losses import CompositeLoss, MultiTaskLoss, AutoTuneMultiTaskLoss from .losses import CompositeLoss, MultiTaskLoss, AutoTuneMultiTaskLoss
from ..network import extract_outputs, extract_labels from ..network import extract_outputs, extract_labels
from ..network.architectures import MonStereoModel from ..network.architectures import LocoModel
from ..utils import set_logger from ..utils import set_logger
@ -35,21 +37,16 @@ class Trainer:
val_task = 'd' val_task = 'd'
lambdas = (1, 1, 1, 1, 1, 1, 1, 1) lambdas = (1, 1, 1, 1, 1, 1, 1, 1)
clusters = ['10', '20', '30', '40'] clusters = ['10', '20', '30', '40']
input_size = dict(mono=34, stereo=68)
output_size = dict(mono=9, stereo=10)
dir_figures = os.path.join('figures', 'losses')
def __init__(self, args): def __init__(self, args):
""" """
Initialize directories, load the data and parameters for the training Initialize directories, load the data and parameters for the training
""" """
# Initialize directories and parameters
dir_out = os.path.join('data', 'models')
if not os.path.exists(dir_out):
warnings.warn("Warning: output directory not found, the model will not be saved")
dir_logs = os.path.join('data', 'logs')
if not os.path.exists(dir_logs):
warnings.warn("Warning: default logs directory not found")
assert os.path.exists(args.joints), "Input file not found" assert os.path.exists(args.joints), "Input file not found"
self.mode = args.mode self.mode = args.mode
self.joints = args.joints self.joints = args.joints
self.num_epochs = args.epochs self.num_epochs = args.epochs
@ -60,10 +57,22 @@ class Trainer:
self.sched_gamma = args.sched_gamma self.sched_gamma = args.sched_gamma
self.hidden_size = args.hidden_size self.hidden_size = args.hidden_size
self.n_stage = args.n_stage self.n_stage = args.n_stage
self.dir_out = dir_out
self.r_seed = args.r_seed self.r_seed = args.r_seed
self.auto_tune_mtl = args.auto_tune_mtl self.auto_tune_mtl = args.auto_tune_mtl
# Select path out
if args.out:
self.path_out = args.out # full path without extension
dir_out, _ = os.path.split(self.path_out)
else:
dir_out = os.path.join('data', 'outputs')
name = 'monoloco_pp' if self.mode == 'mono' else 'monstereo'
now = datetime.datetime.now()
now_time = now.strftime("%Y%m%d-%H%M")[2:]
name_out = name + '-' + now_time + '.pkl'
self.path_out = os.path.join(dir_out, name_out)
assert dir_out, "Directory to save the model not found"
print(self.path_out)
# Select the device # Select the device
use_cuda = torch.cuda.is_available() use_cuda = torch.cuda.is_available()
self.device = torch.device("cuda" if use_cuda else "cpu") self.device = torch.device("cuda" if use_cuda else "cpu")
@ -85,45 +94,28 @@ class Trainer:
self.mt_loss = MultiTaskLoss(losses_tr, losses_val, self.lambdas, self.tasks) self.mt_loss = MultiTaskLoss(losses_tr, losses_val, self.lambdas, self.tasks)
self.mt_loss.to(self.device) self.mt_loss.to(self.device)
if self.mode == 'stereo':
input_size = 68
output_size = 10
else:
input_size = 34
output_size = 9
name = 'monoloco_pp' if self.mode == 'mono' else 'monstereo'
now = datetime.datetime.now()
now_time = now.strftime("%Y%m%d-%H%M")[2:]
name_out = name + '-' + now_time
if not self.no_save:
self.path_model = os.path.join(dir_out, name_out + '.pkl')
self.logger = set_logger(os.path.join(dir_logs, name_out))
self.logger.info( # pylint: disable=logging-fstring-interpolation
f'Training arguments: \ninput_file: {self.joints} \nmode: {self.mode} '
f'\nlearning rate: {args.lr} \nbatch_size: {args.bs}'
f'\nepochs: {args.epochs} \ndropout: {args.dropout} '
f'\nscheduler step: {args.sched_step} \nscheduler gamma: {args.sched_gamma} '
f'\ninput_size: {input_size} \noutput_size: {output_size} \nhidden_size: {args.hidden_size}'
f' \nn_stages: {args.n_stage} \n r_seed: {args.r_seed} \nlambdas: {self.lambdas}'
)
else:
logging.basicConfig(level=logging.INFO)
self.logger = logging.getLogger(__name__)
# Dataloader # Dataloader
self.dataloaders = {phase: DataLoader(KeypointsDataset(self.joints, phase=phase), self.dataloaders = {phase: DataLoader(KeypointsDataset(self.joints, phase=phase),
batch_size=args.bs, shuffle=True) for phase in ['train', 'val']} batch_size=args.bs, shuffle=True) for phase in ['train', 'val']}
self.dataset_sizes = {phase: len(KeypointsDataset(self.joints, phase=phase)) self.dataset_sizes = {phase: len(KeypointsDataset(self.joints, phase=phase))
for phase in ['train', 'val']} for phase in ['train', 'val']}
self.dataset_version = KeypointsDataset(self.joints, phase='train').get_version()
self._set_logger(args)
# Define the model # Define the model
self.logger.info('Sizes of the dataset: {}'.format(self.dataset_sizes)) self.logger.info('Sizes of the dataset: {}'.format(self.dataset_sizes))
print(">>> creating model") print(">>> creating model")
self.model = MonStereoModel(input_size=input_size, output_size=output_size, linear_size=args.hidden_size, self.model = LocoModel(
p_dropout=args.dropout, num_stage=self.n_stage, device=self.device) input_size=self.input_size[self.mode],
output_size=self.output_size[self.mode],
linear_size=args.hidden_size,
p_dropout=args.dropout,
num_stage=self.n_stage,
device=self.device,
)
self.model.to(self.device) self.model.to(self.device)
print(">>> model params: {:.3f}M".format(sum(p.numel() for p in self.model.parameters()) / 1000000.0)) print(">>> model params: {:.3f}M".format(sum(p.numel() for p in self.model.parameters()) / 1000000.0))
print(">>> loss params: {}".format(sum(p.numel() for p in self.mt_loss.parameters()))) print(">>> loss params: {}".format(sum(p.numel() for p in self.mt_loss.parameters())))
@ -158,7 +150,7 @@ class Trainer:
if phase == 'train': if phase == 'train':
self.optimizer.zero_grad() self.optimizer.zero_grad()
outputs = self.model(inputs) outputs = self.model(inputs)
loss, loss_values = self.mt_loss(outputs, labels, phase=phase) loss, _ = self.mt_loss(outputs, labels, phase=phase)
loss.backward() loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 3) torch.nn.utils.clip_grad_norm_(self.model.parameters(), 3)
self.optimizer.step() self.optimizer.step()
@ -188,8 +180,7 @@ class Trainer:
self.logger.info('Best validation Accuracy for {}: {:.3f}'.format(self.val_task, best_acc)) self.logger.info('Best validation Accuracy for {}: {:.3f}'.format(self.val_task, best_acc))
self.logger.info('Saved weights of the model at epoch: {}'.format(best_epoch)) self.logger.info('Saved weights of the model at epoch: {}'.format(best_epoch))
if self.print_loss: self._print_losses(epoch_losses)
print_losses(epoch_losses)
# load best model weights # load best model weights
self.model.load_state_dict(best_model_wts) self.model.load_state_dict(best_model_wts)
@ -255,7 +246,7 @@ class Trainer:
def compute_stats(self, outputs, labels, dic_err, size_eval, clst): def compute_stats(self, outputs, labels, dic_err, size_eval, clst):
"""Compute mean, bi and max of torch tensors""" """Compute mean, bi and max of torch tensors"""
loss, loss_values = self.mt_loss(outputs, labels, phase='val') _, loss_values = self.mt_loss(outputs, labels, phase='val')
rel_frac = outputs.size(0) / size_eval rel_frac = outputs.size(0) / size_eval
tasks = self.tasks[:-1] if self.tasks[-1] == 'aux' else self.tasks # Exclude auxiliary tasks = self.tasks[:-1] if self.tasks[-1] == 'aux' else self.tasks # Exclude auxiliary
@ -333,6 +324,41 @@ class Trainer:
if epoch % 10 == 0: if epoch % 10 == 0:
print(string.format(*format_list)) print(string.format(*format_list))
def _print_losses(self, epoch_losses):
if not self.print_loss:
return
os.makedirs(self.dir_figures, exist_ok=True)
for idx, phase in enumerate(epoch_losses):
for idx_2, el in enumerate(epoch_losses['train']):
plt.figure(idx + idx_2)
plt.title(phase + '_' + el)
plt.xlabel('epochs')
plt.plot(epoch_losses[phase][el][10:], label='{} Loss: {}'.format(phase, el))
plt.savefig(os.path.join(self.dir_figures, '{}_loss_{}.png'.format(phase, el)))
plt.close()
def _set_logger(self, args):
if self.no_save:
logging.basicConfig(level=logging.INFO)
self.logger = logging.getLogger(__name__)
else:
self.path_model = self.path_out
print(self.path_model)
self.logger = set_logger(os.path.splitext(self.path_out)[0]) # remove .pkl
self.logger.info( # pylint: disable=logging-fstring-interpolation
f'\nVERSION: {__version__}\n'
f'\nINPUT_FILE: {args.joints}'
f'\nInput file version: {self.dataset_version}'
f'\nTorch version: {torch.__version__}\n'
f'\nTraining arguments:'
f'\nmode: {self.mode} \nlearning rate: {args.lr} \nbatch_size: {args.bs}'
f'\nepochs: {args.epochs} \ndropout: {args.dropout} '
f'\nscheduler step: {args.sched_step} \nscheduler gamma: {args.sched_gamma} '
f'\ninput_size: {self.input_size[self.mode]} \noutput_size: {self.output_size[self.mode]} '
f'\nhidden_size: {args.hidden_size}'
f' \nn_stages: {args.n_stage} \n r_seed: {args.r_seed} \nlambdas: {self.lambdas}'
)
def debug_plots(inputs, labels): def debug_plots(inputs, labels):
inputs_shoulder = inputs.cpu().numpy()[:, 5] inputs_shoulder = inputs.cpu().numpy()[:, 5]
@ -347,15 +373,6 @@ def debug_plots(inputs, labels):
plt.show() plt.show()
def print_losses(epoch_losses):
for idx, phase in enumerate(epoch_losses):
for idx_2, el in enumerate(epoch_losses['train']):
plt.figure(idx + idx_2)
plt.plot(epoch_losses[phase][el][10:], label='{} Loss: {}'.format(phase, el))
plt.savefig('figures/{}_loss_{}.png'.format(phase, el))
plt.close()
def get_accuracy(outputs, labels): def get_accuracy(outputs, labels):
"""From Binary cross entropy outputs to accuracy""" """From Binary cross entropy outputs to accuracy"""

13
monoloco/utils/__init__.py
View File

@ -1,12 +1,13 @@
from .iou import get_iou_matches, reorder_matches, get_iou_matrix, get_iou_matches_matrix from .iou import get_iou_matches, reorder_matches, get_iou_matrix, get_iou_matches_matrix, get_category, \
from .misc import get_task_error, get_pixel_error, append_cluster, open_annotations, make_new_directory,\ open_annotations
from .misc import get_task_error, get_pixel_error, append_cluster, make_new_directory,\
normalize_hwl, average normalize_hwl, average
from .kitti import check_conditions, get_difficulty, split_training, parse_ground_truth, get_calibration, \ from .kitti import check_conditions, get_difficulty, split_training, get_calibration, \
factory_basename, factory_file, get_category, read_and_rewrite, find_cluster factory_basename, read_and_rewrite, find_cluster
from .camera import xyz_from_distance, get_keypoints, pixel_to_camera, project_3d, open_image, correct_angle,\ from .camera import xyz_from_distance, get_keypoints, pixel_to_camera, project_3d, open_image, correct_angle,\
to_spherical, to_cartesian, back_correct_angles, project_to_pixels to_spherical, to_cartesian, back_correct_angles, project_to_pixels
from .logs import set_logger from .logs import set_logger
from ..utils.nuscenes import select_categories from .nuscenes import select_categories
from ..utils.stereo import mask_joint_disparity, average_locations, extract_stereo_matches, \ from .stereo import mask_joint_disparity, average_locations, extract_stereo_matches, \
verify_stereo, disparity_to_depth verify_stereo, disparity_to_depth

49
monoloco/utils/iou.py
View File

@ -1,4 +1,6 @@
import json
import numpy as np import numpy as np
@ -52,7 +54,7 @@ def get_iou_matches(boxes, boxes_gt, iou_min=0.3):
for idx in indices[::-1]: for idx in indices[::-1]:
box = boxes[idx] box = boxes[idx]
ious = [] ious = []
for idx_gt, box_gt in enumerate(boxes_gt): for box_gt in boxes_gt:
iou = calculate_iou(box, box_gt) iou = calculate_iou(box, box_gt)
ious.append(iou) ious.append(iou)
idx_gt_max = int(np.argmax(ious)) idx_gt_max = int(np.argmax(ious))
@ -96,3 +98,48 @@ def reorder_matches(matches, boxes, mode='left_rigth'):
matches_left = [int(idx) for (idx, _) in matches] matches_left = [int(idx) for (idx, _) in matches]
return [matches[matches_left.index(idx_boxes)] for idx_boxes in ordered_boxes if idx_boxes in matches_left] return [matches[matches_left.index(idx_boxes)] for idx_boxes in ordered_boxes if idx_boxes in matches_left]
def get_category(keypoints, path_byc):
"""Find the category for each of the keypoints"""
dic_byc = open_annotations(path_byc)
boxes_byc = dic_byc['boxes'] if dic_byc else []
boxes_ped = make_lower_boxes(keypoints)
matches = get_matches_bikes(boxes_ped, boxes_byc)
list_byc = [match[0] for match in matches]
categories = [1.0 if idx in list_byc else 0.0 for idx, _ in enumerate(boxes_ped)]
return categories
def get_matches_bikes(boxes_ped, boxes_byc):
matches = get_iou_matches_matrix(boxes_ped, boxes_byc, thresh=0.15)
matches_b = []
for idx, idx_byc in matches:
box_ped = boxes_ped[idx]
box_byc = boxes_byc[idx_byc]
width_ped = box_ped[2] - box_ped[0]
width_byc = box_byc[2] - box_byc[0]
center_ped = (box_ped[2] + box_ped[0]) / 2
center_byc = (box_byc[2] + box_byc[0]) / 2
if abs(center_ped - center_byc) < min(width_ped, width_byc) / 4:
matches_b.append((idx, idx_byc))
return matches_b
def make_lower_boxes(keypoints):
lower_boxes = []
keypoints = np.array(keypoints)
for kps in keypoints:
lower_boxes.append([min(kps[0, 9:]), min(kps[1, 9:]), max(kps[0, 9:]), max(kps[1, 9:])])
return lower_boxes
def open_annotations(path_ann):
try:
with open(path_ann, 'r') as f:
annotations = json.load(f)
except FileNotFoundError:
annotations = []
return annotations

103
monoloco/utils/kitti.py
View File

@ -1,5 +1,4 @@
import math
import os import os
import glob import glob
@ -129,50 +128,6 @@ def split_training(names_gt, path_train, path_val):
return set_train, set_val return set_train, set_val
def parse_ground_truth(path_gt, category, spherical=False, verbose=False):
"""Parse KITTI ground truth files"""
from ..utils import correct_angle, to_spherical
boxes_gt = []
ys = []
truncs_gt = [] # Float from 0 to 1
occs_gt = [] # Either 0,1,2,3 fully visible, partly occluded, largely occluded, unknown
lines = []
with open(path_gt, "r") as f_gt:
for line_gt in f_gt:
line = line_gt.split()
if check_conditions(line_gt, category, method='gt'):
truncs_gt.append(float(line[1]))
occs_gt.append(int(line[2]))
boxes_gt.append([float(x) for x in line[4:8]])
xyz = [float(x) for x in line[11:14]]
hwl = [float(x) for x in line[8:11]]
dd = float(math.sqrt(xyz[0] ** 2 + xyz[1] ** 2 + xyz[2] ** 2))
yaw = float(line[14])
assert - math.pi <= yaw <= math.pi
alpha = float(line[3])
sin, cos, yaw_corr = correct_angle(yaw, xyz)
assert min(abs(-yaw_corr - alpha), (abs(yaw_corr - alpha))) < 0.15, "more than 10 degrees of error"
if spherical:
rtp = to_spherical(xyz)
loc = rtp[1:3] + xyz[2:3] + rtp[0:1] # [theta, psi, z, r]
else:
loc = xyz + [dd]
# cat = 0 if line[0] in ('Pedestrian', 'Person_sitting') else 1
if line[0] in ('Pedestrian', 'Person_sitting'):
cat = 0
else:
cat = 1
output = loc + hwl + [sin, cos, yaw, cat]
ys.append(output)
if verbose:
lines.append(line_gt)
if verbose:
return boxes_gt, ys, truncs_gt, occs_gt, lines
return boxes_gt, ys, truncs_gt, occs_gt
def factory_basename(dir_ann, dir_gt): def factory_basename(dir_ann, dir_gt):
""" Return all the basenames in the annotations folder corresponding to validation images""" """ Return all the basenames in the annotations folder corresponding to validation images"""
@ -191,64 +146,6 @@ def factory_basename(dir_ann, dir_gt):
return set_val return set_val
def factory_file(path_calib, dir_ann, basename, mode='left'):
"""Choose the annotation and the calibration files. Stereo option with ite = 1"""
assert mode in ('left', 'right')
p_left, p_right = get_calibration(path_calib)
if mode == 'left':
kk, tt = p_left[:]
path_ann = os.path.join(dir_ann, basename + '.png.predictions.json')
else:
kk, tt = p_right[:]
path_ann = os.path.join(dir_ann + '_right', basename + '.png.predictions.json')
from ..utils import open_annotations
annotations = open_annotations(path_ann)
return annotations, kk, tt
def get_category(keypoints, path_byc):
"""Find the category for each of the keypoints"""
from ..utils import open_annotations
dic_byc = open_annotations(path_byc)
boxes_byc = dic_byc['boxes'] if dic_byc else []
boxes_ped = make_lower_boxes(keypoints)
matches = get_matches_bikes(boxes_ped, boxes_byc)
list_byc = [match[0] for match in matches]
categories = [1.0 if idx in list_byc else 0.0 for idx, _ in enumerate(boxes_ped)]
return categories
def get_matches_bikes(boxes_ped, boxes_byc):
from . import get_iou_matches_matrix
matches = get_iou_matches_matrix(boxes_ped, boxes_byc, thresh=0.15)
matches_b = []
for idx, idx_byc in matches:
box_ped = boxes_ped[idx]
box_byc = boxes_byc[idx_byc]
width_ped = box_ped[2] - box_ped[0]
width_byc = box_byc[2] - box_byc[0]
center_ped = (box_ped[2] + box_ped[0]) / 2
center_byc = (box_byc[2] + box_byc[0]) / 2
if abs(center_ped - center_byc) < min(width_ped, width_byc) / 4:
matches_b.append((idx, idx_byc))
return matches_b
def make_lower_boxes(keypoints):
lower_boxes = []
keypoints = np.array(keypoints)
for kps in keypoints:
lower_boxes.append([min(kps[0, 9:]), min(kps[1, 9:]), max(kps[0, 9:]), max(kps[1, 9:])])
return lower_boxes
def read_and_rewrite(path_orig, path_new): def read_and_rewrite(path_orig, path_new):
"""Read and write same txt file. If file not found, create open file""" """Read and write same txt file. If file not found, create open file"""
try: try:

10
monoloco/utils/misc.py
View File

@ -1,4 +1,3 @@
import json
import shutil import shutil
import os import os
@ -44,15 +43,6 @@ def get_pixel_error(zz_gt):
return error return error
def open_annotations(path_ann):
try:
with open(path_ann, 'r') as f:
annotations = json.load(f)
except FileNotFoundError:
annotations = []
return annotations
def make_new_directory(dir_out): def make_new_directory(dir_out):
"""Remove the output directory if already exists (avoid residual txt files)""" """Remove the output directory if already exists (avoid residual txt files)"""
if os.path.exists(dir_out): if os.path.exists(dir_out):

23
monoloco/utils/stereo.py
View File

@ -12,7 +12,16 @@ D_MAX = BF / z_min
def extract_stereo_matches(keypoint, keypoints_r, zz, phase='train', seed=0, method=None): def extract_stereo_matches(keypoint, keypoints_r, zz, phase='train', seed=0, method=None):
"""Return binaries representing the match between the pose in the left and the ones in the right""" """
Return:
1) a list of tuples that indicates, for a reference pose in the left image:
- the index of the right pose
- weather the right pose corresponds to the same person as the left pose (stereo match) or not
For example: [(0,0), (1,0), (2,1)] means there are three right poses in the image
and the third one is the same person as the reference pose
2) a flag indicating whether a match has been found
3) number of ambiguous instances, for which is not possible to define whether there is a correspondence
"""
stereo_matches = [] stereo_matches = []
cnt_ambiguous = 0 cnt_ambiguous = 0
@ -70,20 +79,10 @@ def extract_stereo_matches(keypoint, keypoints_r, zz, phase='train', seed=0, met
if idx_matches[num] not in used: if idx_matches[num] not in used:
stereo_matches.append((idx_matches[num], 0)) stereo_matches.append((idx_matches[num], 0))
# elif len(stereo_matches) < 1:
# stereo_matches.append((idx_match, 0))
# Easy-negative
# elif len(idx_matches) > len(stereo_matches):
# stereo_matches.append((idx_matches[-1], 0))
# break # matches are ordered
else: else:
break break
used.append(idx_match) used.append(idx_match)
# Make sure each left has at least a negative match
# if not stereo_matches:
# stereo_matches.append((idx_matches[0], 0))
return stereo_matches, cnt_ambiguous return stereo_matches, cnt_ambiguous
@ -191,7 +190,7 @@ def verify_stereo(zz_stereo, zz_mono, disparity_x, disparity_y):
y_max_difference = (80 / zz_mono) y_max_difference = (80 / zz_mono)
z_max_difference = 1 * zz_mono z_max_difference = 1 * zz_mono
cov = float(np.nanstd(disparity_x) / np.abs(np.nanmean(disparity_x))) # Coefficient of variation cov = float(np.nanstd(disparity_x) / np.abs(np.nanmean(disparity_x))) # pylint: disable=unused-variable
avg_disparity_y = np.nanmedian(disparity_y) avg_disparity_y = np.nanmedian(disparity_y)
return abs(zz_stereo - zz_mono) < z_max_difference and avg_disparity_y < y_max_difference and 1 < zz_stereo < 80 return abs(zz_stereo - zz_mono) < z_max_difference and avg_disparity_y < y_max_difference and 1 < zz_stereo < 80

30
monoloco/visuals/figures.py
View File

@ -7,6 +7,11 @@ import os
import numpy as np import numpy as np
import matplotlib.pyplot as plt import matplotlib.pyplot as plt
from matplotlib.patches import Ellipse from matplotlib.patches import Ellipse
try:
import pandas as pd
DATAFRAME = pd.DataFrame
except ImportError:
DATAFRAME = None
from ..utils import get_task_error, get_pixel_error from ..utils import get_task_error, get_pixel_error
@ -33,7 +38,7 @@ def show_results(dic_stats, clusters, net, dir_fig, show=False, save=False):
excl_clusters = ['all', 'easy', 'moderate', 'hard', '49'] excl_clusters = ['all', 'easy', 'moderate', 'hard', '49']
clusters = [clst for clst in clusters if clst not in excl_clusters] clusters = [clst for clst in clusters if clst not in excl_clusters]
styles = printing_styles(net) styles = printing_styles(net)
for idx_style, style in enumerate(styles.items()): for idx_style in styles:
plt.figure(idx_style, figsize=FIGSIZE) plt.figure(idx_style, figsize=FIGSIZE)
plt.grid(linewidth=GRID_WIDTH) plt.grid(linewidth=GRID_WIDTH)
plt.xlim(x_min, x_max) plt.xlim(x_min, x_max)
@ -183,7 +188,6 @@ def show_method(save, dir_out='data/figures'):
def show_box_plot(dic_errors, clusters, dir_fig, show=False, save=False): def show_box_plot(dic_errors, clusters, dir_fig, show=False, save=False):
import pandas as pd
excl_clusters = ['all', 'easy', 'moderate', 'hard'] excl_clusters = ['all', 'easy', 'moderate', 'hard']
clusters = [int(clst) for clst in clusters if clst not in excl_clusters] clusters = [int(clst) for clst in clusters if clst not in excl_clusters]
methods = ('monstereo', 'pseudo-lidar', '3dop', 'monoloco') methods = ('monstereo', 'pseudo-lidar', '3dop', 'monoloco')
@ -192,7 +196,7 @@ def show_box_plot(dic_errors, clusters, dir_fig, show=False, save=False):
xxs = get_distances(clusters) xxs = get_distances(clusters)
labels = [str(xx) for xx in xxs] labels = [str(xx) for xx in xxs]
for idx, method in enumerate(methods): for idx, method in enumerate(methods):
df = pd.DataFrame([dic_errors[method][str(clst)] for clst in clusters[:-1]]).T df = DATAFRAME([dic_errors[method][str(clst)] for clst in clusters[:-1]]).T
df.columns = labels df.columns = labels
plt.figure(idx, figsize=FIGSIZE) # with 200 dpi it becomes 1920x1440 plt.figure(idx, figsize=FIGSIZE) # with 200 dpi it becomes 1920x1440
@ -289,16 +293,16 @@ def expandgrid(*itrs):
return combinations return combinations
def get_percentile(dist_gmm): # def get_percentile(dist_gmm):
dd_gt = 1000 # dd_gt = 1000
mu_gmm = np.mean(dist_gmm) # mu_gmm = np.mean(dist_gmm)
dist_d = dd_gt * mu_gmm / dist_gmm # dist_d = dd_gt * mu_gmm / dist_gmm
perc_d, _ = np.nanpercentile(dist_d, [18.5, 81.5]) # Laplace bi => 63% # perc_d, _ = np.nanpercentile(dist_d, [18.5, 81.5]) # Laplace bi => 63%
perc_d2, _ = np.nanpercentile(dist_d, [23, 77]) # perc_d2, _ = np.nanpercentile(dist_d, [23, 77])
mu_d = np.mean(dist_d) # mu_d = np.mean(dist_d)
# mm_bi = (mu_d - perc_d) / mu_d # # mm_bi = (mu_d - perc_d) / mu_d
# mm_test = (mu_d - perc_d2) / mu_d # # mm_test = (mu_d - perc_d2) / mu_d
# mad_d = np.mean(np.abs(dist_d - mu_d)) # # mad_d = np.mean(np.abs(dist_d - mu_d))
def printing_styles(net): def printing_styles(net):

14
monoloco/visuals/pifpaf_show.py
View File

@ -1,18 +1,20 @@
# File adapted from https://github.com/vita-epfl/openpifpaf """
Adapted from https://github.com/openpifpaf,
which is: 'Copyright 2019-2021 by Sven Kreiss and contributors. All rights reserved.'
and licensed under GNU AGPLv3
"""
from contextlib import contextmanager from contextlib import contextmanager
import numpy as np import numpy as np
from PIL import Image from PIL import Image
import matplotlib
import matplotlib.pyplot as plt
try: try:
import matplotlib
import matplotlib.pyplot as plt
import scipy.ndimage as ndimage import scipy.ndimage as ndimage
except ImportError: except ImportError:
matplotlib = None
plt = None
ndimage = None ndimage = None
@ -69,7 +71,7 @@ def load_image(path, scale=1.0):
return image return image
class KeypointPainter(object): class KeypointPainter:
def __init__(self, *, def __init__(self, *,
skeleton=None, skeleton=None,
xy_scale=1.0, highlight=None, highlight_invisible=False, xy_scale=1.0, highlight=None, highlight_invisible=False,

7
monoloco/visuals/plot_3d_box.py
View File

@ -10,11 +10,11 @@ def correct_boxes(boxes, hwls, xyzs, yaws, path_calib):
p2_list = [float(xx) for xx in p2_str] p2_list = [float(xx) for xx in p2_str]
P = np.array(p2_list).reshape(3, 4) P = np.array(p2_list).reshape(3, 4)
boxes_new = [] boxes_new = []
for idx, box in enumerate(boxes): for idx in range(boxes):
hwl = hwls[idx] hwl = hwls[idx]
xyz = xyzs[idx] xyz = xyzs[idx]
yaw = yaws[idx] yaw = yaws[idx]
corners_2d, corners_3d = compute_box_3d(hwl, xyz, yaw, P) corners_2d, _ = compute_box_3d(hwl, xyz, yaw, P)
box_new = project_8p_to_4p(corners_2d).reshape(-1).tolist() box_new = project_8p_to_4p(corners_2d).reshape(-1).tolist()
boxes_new.append(box_new) boxes_new.append(box_new)
return boxes_new return boxes_new
@ -58,7 +58,6 @@ def compute_box_3d(hwl, xyz, ry, P):
return corners_2d, np.transpose(corners_3d) return corners_2d, np.transpose(corners_3d)
def roty(t): def roty(t):
""" Rotation about the y-axis. """ """ Rotation about the y-axis. """
c = np.cos(t) c = np.cos(t)
@ -66,7 +65,6 @@ def roty(t):
return np.array([[c, 0, s], [0, 1, 0], [-s, 0, c]]) return np.array([[c, 0, s], [0, 1, 0], [-s, 0, c]])
def project_to_image(pts_3d, P): def project_to_image(pts_3d, P):
""" Project 3d points to image plane. """ Project 3d points to image plane.
Usage: pts_2d = projectToImage(pts_3d, P) Usage: pts_2d = projectToImage(pts_3d, P)
@ -87,7 +85,6 @@ def project_to_image(pts_3d, P):
return pts_2d[:, 0:2] return pts_2d[:, 0:2]
def project_8p_to_4p(pts_2d): def project_8p_to_4p(pts_2d):
x0 = np.min(pts_2d[:, 0]) x0 = np.min(pts_2d[:, 0])
x1 = np.max(pts_2d[:, 0]) x1 = np.max(pts_2d[:, 0])

38
pyproject.toml Normal file
View File

@ -0,0 +1,38 @@
[build-system]
build-backend = "setuptools.build_meta"
requires = ["setuptools", "versioneer-518"]
[tool.pytest.ini_options]
minversion = "6.0"
testpaths = ["tests"]
markers = ["slow: marks tests as slow (deselect with '-m \"not slow\"')"]
[tool.pylint.BASIC]
class-const-naming-style = "snake_case" # checked since pylint 2.7.3
[tool.pylint.IMPORTS]
allow-any-import-level = []
[tool.pylint.SIMILARITIES]
min-similarity-lines = 15 # Minimum lines number of a similarity.
ignore-comments = "yes" # Ignore comments when computing similarities.
ignore-docstrings = "yes" # Ignore docstrings when computing similarities.
ignore-imports = "yes" # Ignore imports when computing similarities.
[tool.pylint.TYPECHECK]
generated-members = ["numpy.*", "torch.*", "cv2.*", "openpifpaf.functional.*"]
ignored-modules = ["openpifpaf.functional"]
disable = [
"missing-docstring",
"too-many-arguments",
"too-many-instance-attributes",
"too-many-locals",
"too-few-public-methods",
"unsubscriptable-object",
"not-callable", # for torch tensors
"invalid-name",
"logging-format-interpolation",
]
[tool.pylint.FORMAT]
max-line-length = 120

13
setup.cfg Normal file
View File

@ -0,0 +1,13 @@
[versioneer]
VCS = git
style = pep440
versionfile_source = monoloco/_version.py
versionfile_build = monoloco/_version.py
tag_prefix = v
#parentdir_prefix =
[pycodestyle]
max-line-length=120
ignore=E731,E741,W503
exclude=monoloco/__init__.py

24
setup.py
View File

@ -1,13 +1,18 @@
from setuptools import setup from setuptools import setup
# extract version from __init__.py # This is needed for versioneer to be importable when building with PEP 517.
with open('monoloco/__init__.py', 'r') as f: # See <https://github.com/warner/python-versioneer/issues/193> and links
VERSION_LINE = [l for l in f if l.startswith('__version__')][0] # therein for more information.
VERSION = VERSION_LINE.split('=')[1].strip()[1:-1]
import os, sys
sys.path.append(os.path.dirname(__file__))
import versioneer
setup( setup(
name='monoloco', name='monoloco',
version=VERSION, version=versioneer.get_version(),
cmdclass=versioneer.get_cmdclass(),
packages=[ packages=[
'monoloco', 'monoloco',
'monoloco.network', 'monoloco.network',
@ -29,15 +34,18 @@ setup(
install_requires=[ install_requires=[
'openpifpaf>=v0.12.1', 'openpifpaf>=v0.12.1',
'matplotlib', 'matplotlib',
'gdown',
], ],
extras_require={ extras_require={
'test': [
'pylint',
'pytest',
'gdown',
'scipy', # for social distancing gaussian blur
],
'eval': [ 'eval': [
'tabulate', 'tabulate',
'sklearn', 'sklearn',
'pandas', 'pandas',
'pylint',
'pytest',
], ],
'prep': [ 'prep': [
'nuscenes-devkit==1.0.2', 'nuscenes-devkit==1.0.2',

1
tests/joints_sample.json
View File

File diff suppressed because one or more lines are too long

1
tests/sample_joints-kitti-mono.json Normal file
View File

File diff suppressed because one or more lines are too long

1
tests/sample_joints-kitti-stereo.json Normal file
View File

File diff suppressed because one or more lines are too long

89
tests/test_iou.ipynb
View File

@ -1,89 +0,0 @@
{
"cells": [
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"import math\n",
"def calculate_iou(box1, box2):\n",
"\n",
" # Calculate the (x1, y1, x2, y2) coordinates of the intersection of box1 and box2. Calculate its Area.\n",
" xi1 = max(box1[0], box2[0])\n",
" yi1 = max(box1[1], box2[1])\n",
" xi2 = min(box1[2], box2[2])\n",
" yi2 = min(box1[3], box2[3])\n",
" inter_area = max((xi2 - xi1), 0) * max((yi2 - yi1), 0) # Max keeps into account not overlapping box\n",
"\n",
" # Calculate the Union area by using Formula: Union(A,B) = A + B - Inter(A,B)\n",
" box1_area = (box1[2] - box1[0]) * (box1[3] - box1[1])\n",
" box2_area = (box2[2] - box2[0]) * (box2[3] - box2[1])\n",
" union_area = box1_area + box2_area - inter_area\n",
"\n",
" # compute the IoU\n",
" iou = inter_area / union_area\n",
"\n",
" return iou"
]
},
{
"cell_type": "code",
"execution_count": 64,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"15.0\n",
"[8.450052369622647, 12.393410142113215, 88.45005236962265, 77.39341014211321]\n",
"0.4850460596873889\n"
]
}
],
"source": [
"x1 = 75\n",
"y1 = 60\n",
"\n",
"box1 = [0, 0, x1, y1]\n",
"alpha = math.atan2(110,75) # good number\n",
"diag = 15\n",
"x_cateto = diag * math.cos(alpha)\n",
"y_cateto = diag * math.sin(alpha)\n",
"print(math.sqrt(x_cateto**2 + y_cateto**2))\n",
"box2 = [x_cateto, y_cateto, x1 + x_cateto + 5, y1 + y_cateto+ 5]\n",
"print(box2)\n",
"print(calculate_iou(box1, box2))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.3"
}
},
"nbformat": 4,
"nbformat_minor": 2
}

69
tests/test_package.py
View File

@ -1,69 +0,0 @@
"""Test if the main modules of the package run correctly"""
import os
import sys
import json
# Python does not consider the current directory to be a package
sys.path.insert(0, os.path.join('..', 'monoloco'))
from PIL import Image
from stereoloco.train import Trainer
from stereoloco.network import MonoLoco
from stereoloco.network.process import preprocess_pifpaf, factory_for_gt
from stereoloco.visuals.printer import Printer
JOINTS = 'tests/joints_sample.json'
PIFPAF_KEYPOINTS = 'tests/002282.png.pifpaf.json'
IMAGE = 'docs/002282.png'
def tst_trainer(joints):
trainer = Trainer(joints=joints, epochs=150, lr=0.01)
_ = trainer.train()
dic_err, model = trainer.evaluate()
return dic_err['val']['all']['mean'], model
def tst_prediction(model, path_keypoints):
with open(path_keypoints, 'r') as f:
pifpaf_out = json.load(f)
kk, _ = factory_for_gt(im_size=[1240, 340])
# Preprocess pifpaf outputs and run monoloco
boxes, keypoints = preprocess_pifpaf(pifpaf_out)
monoloco = MonoLoco(model)
outputs, varss = monoloco.forward(keypoints, kk)
dic_out = monoloco.post_process(outputs, varss, boxes, keypoints, kk)
return dic_out, kk
def tst_printer(dic_out, kk, image_path):
"""Draw a fake figure"""
with open(image_path, 'rb') as f:
pil_image = Image.open(f).convert('RGB')
printer = Printer(image=pil_image, output_path='tests/test_image', kk=kk, output_types=['multi'], z_max=15)
figures, axes = printer.factory_axes()
printer.draw(figures, axes, dic_out, pil_image, save=True)
def test_package():
# Training test
val_acc, model = tst_trainer(JOINTS)
assert val_acc < 2.5
# Prediction test
dic_out, kk = tst_prediction(model, PIFPAF_KEYPOINTS)
assert dic_out['boxes'] and kk
# Visualization test
tst_printer(dic_out, kk, IMAGE)

79
tests/test_train_mono.py Normal file
View File

@ -0,0 +1,79 @@
"""
Adapted from https://github.com/openpifpaf/openpifpaf/blob/main/tests/test_train.py,
which is: 'Copyright 2019-2021 by Sven Kreiss and contributors. All rights reserved.'
and licensed under GNU AGPLv3
"""
import os
import subprocess
import gdown
OPENPIFPAF_MODEL = 'https://drive.google.com/uc?id=1b408ockhh29OLAED8Tysd2yGZOo0N_SQ'
TRAIN_COMMAND = [
'python3', '-m', 'monoloco.run',
'train',
'--joints', 'tests/sample_joints-kitti-mono.json',
'--lr=0.001',
'-e=10',
]
PREDICT_COMMAND = [
'python3', '-m', 'monoloco.run',
'predict',
'docs/002282.png',
'--output_types', 'multi', 'json',
'--decoder-workers=0' # for windows
]
PREDICT_COMMAND_SOCIAL_DISTANCE = [
'python3', '-m', 'monoloco.run',
'predict',
'docs/frame0032.jpg',
'--social_distance',
'--output_types', 'front', 'bird',
'--decoder-workers=0' # for windows'
]
def test_train_mono(tmp_path):
# train a model
train_cmd = TRAIN_COMMAND + ['--out={}'.format(os.path.join(tmp_path, 'train_test.pkl'))]
print(' '.join(train_cmd))
subprocess.run(train_cmd, check=True, capture_output=True)
print(os.listdir(tmp_path))
# find the trained model checkpoint and download pifpaf one
final_model = next(iter(f for f in os.listdir(tmp_path) if f.endswith('.pkl')))
pifpaf_model = os.path.join(tmp_path, 'pifpaf_model.pkl')
print('Downloading OpenPifPaf model in temporary folder')
gdown.download(OPENPIFPAF_MODEL, pifpaf_model)
# run predictions with that model
model = os.path.join(tmp_path, final_model)
print(model)
predict_cmd = PREDICT_COMMAND + [
'--model={}'.format(model),
'--checkpoint={}'.format(pifpaf_model),
'-o={}'.format(tmp_path),
]
print(' '.join(predict_cmd))
subprocess.run(predict_cmd, check=True, capture_output=True)
print(os.listdir(tmp_path))
assert 'out_002282.png.multi.png' in os.listdir(tmp_path)
assert 'out_002282.png.monoloco.json' in os.listdir(tmp_path)
predict_cmd_sd = PREDICT_COMMAND_SOCIAL_DISTANCE + [
'--model={}'.format(model),
'--checkpoint={}'.format(pifpaf_model),
'-o={}'.format(tmp_path),
]
print(' '.join(predict_cmd_sd))
subprocess.run(predict_cmd_sd, check=True, capture_output=True)
print(os.listdir(tmp_path))
assert 'out_frame0032.jpg.front.png' in os.listdir(tmp_path)
assert 'out_frame0032.jpg.bird.png' in os.listdir(tmp_path)

59
tests/test_train_stereo.py Normal file
View File

@ -0,0 +1,59 @@
"""
Adapted from https://github.com/openpifpaf/openpifpaf/blob/main/tests/test_train.py,
which is: 'Copyright 2019-2021 by Sven Kreiss and contributors. All rights reserved.'
and licensed under GNU AGPLv3
"""
import os
import subprocess
import gdown
OPENPIFPAF_MODEL = 'https://drive.google.com/uc?id=1b408ockhh29OLAED8Tysd2yGZOo0N_SQ'
TRAIN_COMMAND = [
'python3', '-m', 'monoloco.run',
'train',
'--mode=stereo',
'--joints', 'tests/sample_joints-kitti-stereo.json',
'--lr=0.001',
'-e=20',
]
PREDICT_COMMAND = [
'python3', '-m', 'monoloco.run',
'predict',
'--mode=stereo',
'--glob', 'docs/000840*.png',
'--output_types', 'multi', 'json',
'--decoder-workers=0', # for windows'
]
def test_train_stereo(tmp_path):
# train a model
train_cmd = TRAIN_COMMAND + ['--out={}'.format(os.path.join(tmp_path, 'train_test.pkl'))]
print(' '.join(train_cmd))
subprocess.run(train_cmd, check=True, capture_output=True)
print(os.listdir(tmp_path))
# find the trained model checkpoint
final_model = next(iter(f for f in os.listdir(tmp_path) if f.endswith('.pkl')))
pifpaf_model = os.path.join(tmp_path, 'pifpaf_model.pkl')
print('Downloading OpenPifPaf model in temporary folder')
gdown.download(OPENPIFPAF_MODEL, pifpaf_model)
# run predictions with that model
model = os.path.join(tmp_path, final_model)
predict_cmd = PREDICT_COMMAND + [
'--model={}'.format(model),
'--checkpoint={}'.format(pifpaf_model),
'-o={}'.format(tmp_path),
]
print(' '.join(predict_cmd))
subprocess.run(predict_cmd, check=True, capture_output=True)
print(os.listdir(tmp_path))
assert 'out_000840.png.multi.png' in os.listdir(tmp_path)

5
tests/test_utils.py
View File

@ -1,12 +1,14 @@
import os import os
import sys import sys
from monoloco.utils import pixel_to_camera
from monoloco.utils import get_iou_matrix
# Python does not consider the current directory to be a package # Python does not consider the current directory to be a package
sys.path.insert(0, os.path.join('..', 'monoloco')) sys.path.insert(0, os.path.join('..', 'monoloco'))
def test_iou(): def test_iou():
from stereoloco.utils import get_iou_matrix
boxes_pred = [[1, 100, 1, 200]] boxes_pred = [[1, 100, 1, 200]]
boxes_gt = [[100., 120., 150., 160.],[12, 110, 130., 160.]] boxes_gt = [[100., 120., 150., 160.],[12, 110, 130., 160.]]
iou_matrix = get_iou_matrix(boxes_pred, boxes_gt) iou_matrix = get_iou_matrix(boxes_pred, boxes_gt)
@ -14,7 +16,6 @@ def test_iou():
def test_pixel_to_camera(): def test_pixel_to_camera():
from stereoloco.utils import pixel_to_camera
kk = [[718.3351, 0., 600.3891], [0., 718.3351, 181.5122], [0., 0., 1.]] kk = [[718.3351, 0., 600.3891], [0., 718.3351, 181.5122], [0., 0., 1.]]
zz = 10 zz = 10
uv_vector = [1000., 400.] uv_vector = [1000., 400.]

23
tests/test_visuals.py
View File

@ -1,23 +0,0 @@
import os
import sys
from collections import defaultdict
from PIL import Image
# Python does not consider the current directory to be a package
sys.path.insert(0, os.path.join('..', 'monoloco'))
def test_printer():
"""Draw a fake figure"""
from stereoloco.visuals.printer import Printer
test_list = [[718.3351, 0., 600.3891], [0., 718.3351, 181.5122], [0., 0., 1.]]
boxes = [xx + [0] for xx in test_list]
kk = test_list
dict_ann = defaultdict(lambda: [1., 2., 3.], xyz_real=test_list, xyz_pred=test_list, uv_shoulders=test_list,
boxes=boxes, boxes_gt=boxes)
with open('docs/002282.png', 'rb') as f:
pil_image = Image.open(f).convert('RGB')
printer = Printer(image=pil_image, output_path=None, kk=kk, output_types=['combined'])
figures, axes = printer.factory_axes()
printer.draw(figures, axes, dict_ann, pil_image)

1855
versioneer.py Normal file
View File

File diff suppressed because it is too large Load Diff