Action Recognition

Setup

1. Download HMDB51 dataset

   wget http://ftp.tugraz.at/pub/feichtenhofer/tsfusion/data/hmdb51_jpegs_256.zip
   wget http://ftp.tugraz.at/pub/feichtenhofer/tsfusion/data/hmdb51_tvl1_flow.zip
   
   mv hmdb51_jpegs_256.zip ar/data
   mv hmdb51_tvl1_flow.zip ar/data
   
   cd ar/data
   
   unzip hmdb51_jpegs_256.zip
   unzip hmdb51_tvl1_flow.zip

2. Install requriements:

    conda create --name ar python=3.7
    conda activate ar
    conda install opencv

Train

RGB

python rgb_train.py

Flow

python motion_train.py

Two Stream

python multi_stream_train.py --streams=rgb,flow

• If use pretrained model and want to freeze the parameters of the pretrained models, run

  python multi_stream_train.py --streams=rgb,flow --use_pretrained=True --freeze_stream_models=True

  If use pretrained model without freezing the parameters, note that the new models will replace the pretrained models after training.

• multi_stream_add.py is for summing predictions from the pretrained models of all streams, run

  python multi_stream_add.py --streams=rgb,flow --use_pretrained=True --freeze_stream_models=True

  This program will print train / val acc.

References

The code is borrowed from or inspired by:

• https://github.com/artest08/LateTemporalModeling3DCNN
• https://github.com/jfzhang95/pytorch-video-recognition