AtGCN

A lightweight Spatiotemporal GCN for Ataxia Detection. Official implementation for AtGCN, accepted to AIME-2025 as a full length paper + talk.

To make our pipeline fuly reproducible, we provide all the code used and also instructions on how to reproduce.

Data Preparation - V1

Please download the first dataset we have used from here; if you want to use your own dataset, the same instructions as below follow.

1. First create a folder named data/ in the root directory of the project.
2. Use the notebooks, FRCNN.ipynb, SORT.ipynb and finally OpenPose.ipynb in this order from the src/preprocess/ directory on Google Colab to get keypoints of all patients, store these inside the data/ directory in a folder named final_keypoints/. (Note: this step takes ~4 hours for the 149 videos we had using a T4 GPU)
3. Using these keypoints we can now extract non-overlapping gait cycles. To do this, run the gait_extractor.py script from the src/preprocess/ directory. This will create a folder named non_overlapping_gait_cycles/ inside the data/ directory (or gait_cycles/ if you want to extract overlapping gait cycles, we use non-overlapping gait-cycles in the paper and have not tested the overlapping case). Run with -

python src/preprocess/gait_extractor.py 1

This should print something like -

Minimum number of peaks: xx at yy
Average number of peaks: xy.z
Maximum number of peaks: zz
Minimum number of gait cycles: xx at yy
Average number of gait cycles: xy.z
Maximum number of gait cycles: zz
Number of gait cycles: abc
Minimum length of gait cycle: xx at yy
Average length of gait cycle: xy.z
Maximum length of gait cycle: zz at yy

4. Finally, we create CSV files from which we can quickly retrieve data and use it in our training loop, you can use create_csvs.py from the src/preprocess/ directory to create these CSV files, in our format.

This will create the folder with all the files and extracted Gait Cycles (overlapping and non-overlapping, can do both).

Data Preparation - V2

We have another dataset of 40 videos, which can be downloaded from here / paper. This already has extracted keypoints. NOTE: according to current preprocessing, you can only use one dataset at a time, as the CSV files and folders are overwritten. To switch between datasets, just rename the folders accordingly.

1. Use the create_npy.py from the src/preprocess/ directory, which will store the keypoints in .npy files and create a V2.csv file which contains a mapping of original files to assigned IDs and their labels, we also add a center coordinate to the keypoints (last keypoint).
2. Then use the gait_extractor.py to extract non-overlapping gait cycles (will be saved similarly to the first dataset). Run with

python src/preprocess/gait_extractor.py 2

3. Finally, create the CSV files using create_csvs.py from the src/preprocess/ directory.

A few notes on this dataset:

• The dataset available online is the augmented version, we deaugment and use it in our case.
• The authors report 31 augmentations + 1 original, however, their implementation and the trend in the dataset observed by plotting shows original + 32 augmentations, so we divide the dataset by 33 and use the first [:x//33] frames per video.

Model checkpoints

We have already uploaded the required checkpoints inside ckpts/. These were taken from -

• STGCN checkpoints : here.
• GaitGraph checkpoints : here.

Training

First we need to install the dependencies (We have used python 3.9.19)

pip install -r requirements.txt

OR if you use conda, then first run

conda create -n ataxia-sync python=3.9.19

To train a model you can use the runner.sh with sbatch on an HPC or with bash on your local machine.

bash runner.sh

You can also check the flags using

python src/trainer.py --help # or -h

To reproduce our results, you can run the runner.sh as is. A full 10-fold CV run takes about 40 minutes on a single V100 GPU.

Analysis

You can also reproduce all the plots and tables in the paper using src/analyze.py, just uncomment all the lines after if __name__ == "__main__", and run,

python3 src/analyze.py

Main results

Please create an issue if you need some functionality or the code doesn't work as intended. Thank you!

Citation

We will update this with a newer version soon, for now please use the following:

@misc{bania2024atgcn,
      title={AtGCN: A Graph Convolutional Network For Ataxic Gait Detection}, 
      author={Karan Bania and Tanmay Verlekar},
      year={2024},
      eprint={2410.22862},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.22862}, 
}