Custom Training Skeleton based Action Recognition Models

[rahul1728jha]

Hi All,

I am looking forward to train the skeleton based models for action recognition for custom dataset.
However i have couple of questions :

1. There are 4 models supported right now, AGCN, PoseC3D, STGCN and STGCN++.
   But the custom training is available only for PoseC3D. I understand that to do a custom training first the key points should be
   prepared based on the link :
   https://github.com/open-mmlab/mmaction2/blob/main/configs/skeleton/posec3d/custom_dataset_training.md
   Do the other models not support custom training, if yes can someone please share the link for that
2. What should be the data collection format as in :
   Assuming an activity as talking on phone
   Video collected for 10 seconds where talking on phone is happening at t=4 to t=6
   Should the final video for whom pose has to be extracted be only clip of t=4 to t=6?
   Should there be only 1 person in the frame : If there are 2 people in the video, one is talking on phone, other is not doing anything, is it a valid scenario or all the data collected should be 1 person in the frame.
3. Does the data collected need to be in NTU format or Kinetics format or any other format?
4. Do the poses need to be extracted through the script https://github.com/open-mmlab/mmaction2/blob/90fc8440961987b7fe3ee99109e2c633c4e30158/tools/data/skeleton/ntu_pose_extraction.py work or any pose estimation model can be used