menorashid/action_graphs

Paper

This repository provides code for:

Action Graphs: Weakly-supervised Action Localization with Graph Convolution Networks. Maheen Rashid, Hedvig Kjellström, Yong Jae Lee. WACV 2020.

If you find this repo useful please cite our work:

@inproceedings{rashid2020action,
  title={Action Graphs: Weakly-supervised Action Localization with Graph Convolution Networks},
  author={Rashid, Maheen and Kjellstr{\"o}m, Hedvig and Lee, Yong Jae},
  booktitle={Winter Conference on Applications of Computer Vision},
  year={2020}
}

For questions contact Maheen Rashid (mhnrashid at ucdavis dot edu)

Getting Started

Download the code from GitHub:

git clone https://github.com/menorashid/action_graphs.git
cd action_graphs

This repo requires Python 2.7.

It'll be easiest to set up a virtual environment for this repo.

pip install virtualenv
virtualenv --python=/usr/bin/python2.7 <path/to/action_graphs_venv>

Activate virtual env and install all needed requirements

source <path/to/action_graphs_venv>/bin/activate
cd <path/to/action_graph_git_repo>
pip install -r requirements.txt

Dataset

Download features extracted from Kinetrics pretrained i3D

UCF101 Features (Dropbox link) (Google Drive link) (5.5 GB)

ActivityNet Features (Dropbox link) (Google Drive link) (8.6 GB)

Charades Features (Will be uploaded soon)

Run the following commands

cd data
tar -xzvf <path to data tar file>

Thank you so much to Sujoy Paul for originally sharing the UCF101 and ActivityNet features here

Models

To download the pretrained models go here for Dropbox and here for Google Drive (90 MB)

Run the following commands

cd experiments
tar -xzvf <path to models tar file>

Testing

To test pretrained model run the following line after uncommenting the relevant lines in main

cd code
python -W ignore test_pretrained.py

Results would be printed and logged in log.txt file in a results folder in the relevant model folder

experiments/action_graphs_ucf/results_model_249_0_-0.9_-2/log.txt

Training

Script for training the full model with varying values of d as shown in Table 4 in the paper (uncomment relevant lines in main):

cd code
python -W ignore exps_deno.py

Script for training the models from Table 2 and Table 3 in the paper (uncomment relevant lines in main):

cd code
python -W ignore exps_ablation_etc.py

This will save and log training in the following dir

experiments/<model name>/<model params>/<training params>/

The output files would be similar to the pretrained model data shared above and include a .txt log file, .png graphs with losses and accuracy, and .pt model files.

Please email any questions to me (mhnrashid at ucdavis dot edu). I may not get around to looking at Git issues. Thank you for looking at our paper and code!

menorashid / action_graphs