evaluation on bottom center

src/eval/generate_kitti.py

@@ -53,6 +53,9 @@ def generate_kitti(model, dir_ann, p_dropout=0.2, n_dropout=0):
             outputs, varss = monoloco.forward(keypoints, kk)
             dds_geom = eval_geometric(keypoints, kk, average_y=0.48)
 
+        if basename == '001782':
+            aa = 5
+
         # Save the file
         all_outputs = [outputs.detach().cpu(), varss.detach().cpu(), dds_geom]
         all_inputs = [boxes, keypoints]
@@ -75,7 +78,7 @@ def save_txts(path_txt, all_inputs, all_outputs, all_params):
     uv_boxes, keypoints = all_inputs[:]
     kk, tt = all_params[:]
 
-    uv_centers = get_keypoints(keypoints, mode='center')
+    uv_centers = get_keypoints(keypoints, mode='bottom')  # Kitti uses the bottom center to calculate depth
     xy_centers = pixel_to_camera(uv_centers, kk, 1)
     zzs = xyz_from_distance(outputs[:, 0:1], xy_centers)[:, 2].tolist()
 

src/models/trainer.py

@@ -216,7 +216,6 @@ class Trainer:
 
         # To load a model instead of using the trained one
         if load:
-            # self.path_out = os.path.join(self.dir_out, 'best_model_paper.pickle')
             self.model.load_state_dict(torch.load(model, map_location=lambda storage, loc: storage))
 
         # Average distance on training and test set after unnormalizing

src/utils/camera.py

@@ -76,7 +76,7 @@ def get_keypoints(keypoints, mode):
         keypoints = keypoints.unsqueeze(0)
 
     assert len(keypoints.size()) == 3 and keypoints.size()[1] == 3, "tensor dimensions not recognized"
-    assert mode in ['center', 'head', 'shoulder', 'hip' , 'ankle']
+    assert mode in ['center', 'bottom', 'head', 'shoulder', 'hip', 'ankle']
 
     kps_in = keypoints[:, 0:2, :]  # (m, 2, 17)
     if mode == 'center':
@@ -84,6 +84,13 @@ def get_keypoints(keypoints, mode):
         kps_min, _ = kps_in.min(2)
         kps_out = (kps_max - kps_min) / 2 + kps_min   # (m, 2) as keepdims is False
 
+    elif mode == 'bottom':  # bottom center for kitti evaluation
+        kps_max, _ = kps_in.max(2)
+        kps_min, _ = kps_in.min(2)
+        kps_out_x = (kps_max[:, 0:1] - kps_min[:, 0:1]) / 2 + kps_min[:, 0:1]
+        kps_out_y = kps_max[:, 1:2]
+        kps_out = torch.cat((kps_out_x, kps_out_y), -1)
+
     elif mode == 'head':
         kps_out = kps_in[:, :, 0:5].mean(2)