Error with pose estimation models on videos with multiple people

[leeping-ng]

Hi guys, I encountered the following error when using MoveNet (multipose) and HRNet on videos with multiple people:

2023-08-20 12:05:49 root  ERROR:  Traceback (most recent call last):
   File "demo.py", line 146, in <module>
    pose_pipeline(args.video_path, args.save_dir)
   File "demo.py", line 99, in pose_pipeline
    runner.run()
   File "/home/leeping/miniconda3/envs/pkd/lib/python3.8/site-packages/peekingduck/runner.py", line 127, in run
    outputs = node.run(inputs)
   File "/home/leeping/miniconda3/envs/pkd/lib/python3.8/site-packages/peekingduck/pipeline/nodes/model/hrnet.py", line 74, in run
    keypoints, keypoint_scores, keypoint_conns = self.model.predict(
   File "/home/leeping/miniconda3/envs/pkd/lib/python3.8/site-packages/peekingduck/pipeline/nodes/model/hrnetv1/hrnet_model.py", line 65, in predict
    return self.detector.predict(frame, detected_bboxes)
   File "/home/leeping/miniconda3/envs/pkd/lib/python3.8/site-packages/peekingduck/pipeline/nodes/model/hrnetv1/hrnet_files/detector.py", line 81, in predict
    poses, keypoint_scores, keypoint_conns = self._postprocess(
   File "/home/leeping/miniconda3/envs/pkd/lib/python3.8/site-packages/peekingduck/pipeline/nodes/model/hrnetv1/hrnet_files/detector.py", line 144, in _postprocess
    keypoint_conns = get_keypoint_conns(normalized_keypoints, kp_masks)
   File "/home/leeping/miniconda3/envs/pkd/lib/python3.8/site-packages/peekingduck/pipeline/nodes/model/hrnetv1/hrnet_files/postprocessing.py", line 126, in get_keypoint_conns
    return np.array(compiled_connections)
 
2023-08-20 12:05:49 root  ERROR:  ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (15,) + inhomogeneous part.

From Stack Overflow, the root cause seems to be the numpy version. I encountered the crash with numpy==1.24.4 when I did a fresh install of PeekingDuck, and upon downgrading to numpy==1.21.1, the issue was fixed. Perhaps you can consider adding an upper limit to the numpy version numpy >= 1.17.3, <=1.21.2 in requirements.txt. Cheers!