search

open-mmlab / mmaction2

OpenMMLab's Next Generation Video Understanding Toolbox and Benchmark
https://mmaction2.readthedocs.io
Apache License 2.0
4.28k stars 1.25k forks source link

Spatial Temporal Action Detection SlowFast with my data. #1002

Closed mostar39 closed 3 years ago

mostar39 commented 3 years ago

I am learning Spatial Temporal Action Detection SlowFast with my data. However, in the case of AVA data, I saw that the video length was fixed from 902 to 1798. I would like to proceed with learning with each video being very short, but the length of each video is different. I changed timestamp_start in ava_dataset.py to 1 according to my data, can I change it like this? Also, I don't know how to set timstamp_end because the length of the video is different. Is there any problem with learning even if the length of the video is different?

mostar39 commented 3 years ago

And I wonder. written in the config file dict(type='SampleAVAFrames', clip_len=32, frame_interval=2), What is the meaning of My data also cut a video into a photo at 30fps like AVA data, is clip_len meaningful with that?

kennymckormick commented 3 years ago
  1. It's OK to have videos of different lengths. However, you may need to modify the AVADataset.
  2. clip_len means the roi_feature of the bbox is extracted from a clip with clip_len x frame_interval frames. So keep its original value.