open-mmlab / mmaction2

OpenMMLab's Next Generation Video Understanding Toolbox and Benchmark
https://mmaction2.readthedocs.io
Apache License 2.0
4.31k stars 1.25k forks source link

Impossibility to understand how to create and train custom dataset[Docs] #2839

Open gguzzy opened 6 months ago

gguzzy commented 6 months ago

The doc issue

Good morning everyone, I want to point out that the 'ntu_pose_extraction,py' is not working as expected, morever for me it just works on cpu device.

Moreover, I would like to understand, how starting from raw videos we can obtain spatio temporal annotated videos to put as input features the skeleton actions to feed both 2d till 4d stream (as wrote in the paper).

My question is, how would we be able to convert these video into feedable annotations? Creating pickles seems a bit weird, since it is reported that each one should be merged together to train our final model, but with which criteria?

An example on how to create a small dataset and train it wouldnt be bad at all, if you want I might help you building that for further references.

Thanks a lot.

Gianluca

Suggest a potential alternative/fix

No response