search

kennymckormick / pyskl

A toolbox for skeleton-based action recognition.
Apache License 2.0
911 stars 175 forks source link

Custom Training for PoseC3D and STGCN++ #195

Open rahul1728jha opened 11 months ago

rahul1728jha commented 11 months ago

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 2 models supported right now, PoseC3D, STGCN++.
  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. Once a final skeleton dataset is prepared, will it work for both the models i.e. PoseC3D and STGCN++
dqj5182 commented 2 months ago

Same question