DECOMPL: Decompositional Learning with Attention Pooling for Group Activity Recognition from a Single Volleyball Image

This is an official implementation of DECOMPL: Decompositional Learning with Attention Pooling for Group Activity Recognition from a Single Volleyball Image. In this repository, we provide PyTorch code for training and testing as described in the paper.

Preparing Dataset

1. Download VD from following link: Volleyball dataset.
2. Unzip the dataset (~60 GB) file into a directory named data and set its name to volleyball_videos
3. Download the file tracks_normalized.pkl from cvlab-epfl/social-scene-understanding and put it into the directory data
4. Finally, place DECOMPL and our reannotations on the same directory

Getting Started

1. Conda (Recommended):

   conda create -n DECOMPL
   conda activate DECOMPL

2. Pip

   pip install -r requirements.txt

3. Training and Validation: Modify test_only argument in /scripts/run_model_volleyball.py to train or validate. To use the pretrained weights set load_path to checkpoint_weights_volleyball_half.pth

   cd PROJECT_PATH 
   python scripts/run_model.py

4. Additionally, Training and Validation for Collective Activity Dataset: Follow the same instructions as for the Volleyball Dataset on the scripts /scripts/run_model_collective.py and pretrained weights checkpoint_weights_collective_half.pth

Note: The weights provided are converted to half precision due to size constraints.

@article{demirel2023decompl,
  title={DECOMPL: Decompositional Learning with Attention Pooling for Group Activity Recognition from a Single Volleyball Image},
  author={Demirel, Berker and Ozkan, Huseyin},
  journal={arXiv preprint arXiv:2303.06439},
  year={2023}
}