fix: silu pose mismatch in oumvlp pose pkl extraction

datasets/pretreatment.py

@@ -3,6 +3,7 @@ import argparse
 import logging
 import multiprocessing as mp
 import os
+import re
 import pickle
 from collections import defaultdict
 from functools import partial
@@ -127,7 +128,7 @@ def pretreat(input_path: Path, output_path: Path, img_size: int = 64, workers: i
             progress.update(1)
     logging.info('Done')
 
-def txts2pickle(txt_groups: Tuple, output_path: Path, verbose: bool = False, dataset='CASIAB') -> None:
+def txts2pickle(txt_groups: Tuple, output_path: Path, verbose: bool = False, dataset='CASIAB', **kwargs) -> None:
     """
     Reads a group of images and saves the data in pickle format.
 
@@ -137,18 +138,50 @@ def txts2pickle(txt_groups: Tuple, output_path: Path, verbose: bool = False, dat
         img_size (int, optional): Image resizing size. Defaults to 64.
         verbose (bool, optional): Display debug info. Defaults to False.
     """    
+    def pose_silu_match_score(pose: np.ndarray, silu: np.ndarray):
+        pose_coord = pose[:,:2].astype(np.int32)
+        
+        H, W, *_ = silu.shape
+        valid_joints = (pose_coord[:, 1] >=0) & (pose_coord[:, 1] < H) & \
+                       (pose_coord[:, 0] >=0) & (pose_coord[:, 0] < W)
+        if np.sum(valid_joints) == len(pose_coord):
+            # only calculate score for points that are inside the silu img
+            # use the sum of all joints' pixel intensity as the score
+            return np.sum(silu[pose_coord[:, 1], pose_coord[:, 0]])
+        else:
+            # if pose coord is out of bound, return -inf
+            return -np.inf
 
     sinfo = txt_groups[0]
     txt_paths = txt_groups[1]
     to_pickle = []
     if dataset == 'OUMVLP':
+        oumvlp_rearrange_silu_path = kwargs.get('oumvlp_rearrange_silu_path', None)
         for txt_file in sorted(txt_paths):
             try:
                 with open(txt_file) as f:
                     jsondata = json.load(f)
-                if len(jsondata['people'])==0:
+                person_num = len(jsondata['people'])
+                if person_num==0:
                     continue
+                elif person_num == 1:
                     data = np.array(jsondata["people"][0]["pose_keypoints_2d"]).reshape(-1,3)
+                else:
+                    # load the reference silu image
+                    img_name = re.findall(r'\d{4}', os.path.basename(txt_file))[-1] + '.png'
+                    img_path = os.path.join(oumvlp_rearrange_silu_path, *sinfo, img_name)
+                    if not os.path.exists(img_path):
+                        logging.warning(f'Pose reference silu({img_path}) not exists.')
+                        continue
+                    silu_img = cv2.imread(img_path, cv2.IMREAD_GRAYSCALE)
+                    
+                    # determine which pose has the highest matching score
+                    person_poses = [np.array(p["pose_keypoints_2d"]).reshape(-1,3) for p in jsondata['people']]
+                    max_score_idx = np.argmax([pose_silu_match_score(p, silu_img) for p in person_poses])
+                    
+                    # use the pose with the highest matching score to be the pkl data
+                    data = person_poses[max_score_idx]
+                    
                 to_pickle.append(data)
             except:
                 print(txt_file)
@@ -174,7 +207,7 @@ def txts2pickle(txt_groups: Tuple, output_path: Path, verbose: bool = False, dat
 
 
 
-def pretreat_pose(input_path: Path, output_path: Path, workers: int = 4, verbose: bool = False, dataset='CASIAB') -> None:
+def pretreat_pose(input_path: Path, output_path: Path, workers: int = 4, verbose: bool = False, dataset='CASIAB', **kwargs) -> None:
     """Reads a dataset and saves the data in pickle format.
 
     Args:
@@ -208,7 +241,10 @@ def pretreat_pose(input_path: Path, output_path: Path, workers: int = 4, verbose
 
     with mp.Pool(workers) as pool:
         logging.info(f'Start pretreating {input_path}')
-        for _ in pool.imap_unordered(partial(txts2pickle, output_path=output_path, verbose=verbose, dataset=args.dataset), txt_groups.items()):
+        for _ in pool.imap_unordered(
+            partial(txts2pickle, output_path=output_path, verbose=verbose, dataset=args.dataset, **kwargs), 
+            txt_groups.items()
+        ):
             progress.update(1)
     logging.info('Done')
 
@@ -224,6 +260,8 @@ if __name__ == '__main__':
     parser.add_argument('-d', '--dataset', default='CASIAB', type=str, help='Dataset for pretreatment.')
     parser.add_argument('-v', '--verbose', default=False, action='store_true', help='Display debug info.')
     parser.add_argument('-p', '--pose', default=False, action='store_true', help='Processing pose.')
+    parser.add_argument('--oumvlp_rearrange_silu_path', default='', type=str, 
+                        help='Root path of the rearranged oumvlp silu dataset. This argument is only used in extracting oumvlp pose pkl.')
     args = parser.parse_args()
 
     logging.basicConfig(level=logging.INFO, filename=args.log_file, filemode='w', format='[%(asctime)s - %(levelname)s]: %(message)s')
@@ -234,6 +272,15 @@ if __name__ == '__main__':
         for k, v in args.__dict__.items():
             logging.debug(f'{k}: {v}')
     if args.pose:
-        pretreat_pose(input_path=Path(args.input_path), output_path=Path(args.output_path), workers=args.n_workers, verbose=args.verbose, dataset=args.dataset)
+        if args.dataset.lower() == "oumvlp":
+            assert args.oumvlp_rearrange_silu_path, "Please specify the path to the rearranged OUMVLP dataset using `--oumvlp_rearrange_silu_path` argument."
+        pretreat_pose(
+            input_path=Path(args.input_path), 
+            output_path=Path(args.output_path), 
+            workers=args.n_workers, 
+            verbose=args.verbose, 
+            dataset=args.dataset,
+            oumvlp_rearrange_silu_path=os.path.abspath(args.oumvlp_rearrange_silu_path)
+        )
     else:
         pretreat(input_path=Path(args.input_path), output_path=Path(args.output_path), img_size=args.img_size, workers=args.n_workers, verbose=args.verbose, dataset=args.dataset)