Custom Dataset Training with PoseC3D problem

[Disisxhy]

Once I skip the 'ntu_det_postproc' step as noticed and then run 'python ntu_pose_extraction.py S001C001P001R001A001_rgb.avi S001C001P001R001A001_rgb.pkl', it shows the error below:

Hope somebody can help

[kennymckormick]

Hello, in PR #1295, I add an option named 'skip_postproc'. Please try to extract pose with this flag, and check if it can resolve your issues.

[Disisxhy]

the ‘skip_postproc' doesn't work well. comes the error 'AttributeError: 'abstractproperty' object has no attribute 'skip_postproc'

since the 'skip' is not necessary for me, I simply change ' num_frame, num_person ' thing and it woks perfectly! thx so much!

[bit-scientist]

Hi, @Disisxhy, basically you don't exactly have to use ntu_pose_extraction.py as it is a burden to switch off ntu_pose_extraction function, you can modify demo_posec3d.py a bit to extract necessary keys from fake_anno (in demo_posec3d.py) and dump into a pickle file. Note that you may have to use your video's name for frame_dirs value and, its string split for labels value (if the video name contains class label ) as shown at #1258. Hope that helps.

[shelly-xue]

@bit-scientist @kennymckormick hi ,my custom videos have two people, if i use ntu_pose_extraction.py to extract poses, Can it be used directly for training? because we need to maintain the continuity of everyone's time, am i right?

[kennymckormick]

I think it can be directly used for PoseC3D training, but maybe not ST-GCN. Since for ST-GCN, the skeleton input should be several continuous sequences, while ntu_pose_extraction.py does not guarantee this: the person 1 in frame t and the person 1 in frame t + 1 may not be the same person.

@bit-scientist @kennymckormick hi ,my custom videos have two people, if i use ntu_pose_extraction.py to extract poses, Can it be used directly for training? because we need to maintain the continuity of everyone's time, am i right?