AG-MAE

AG-MAE: Anatomically Guided Spatio-Temporal Masked Auto-Encoder for Online Hand Gesture Recognition

Updates

• Trained models available soon !

Installation

• Create and activate conda environment:

conda create -n agmae python=3.9
conda activate agmae

• Install all dependencies:

pip install -r requirements.txt

Dataset

Download the SHREC'21 dataset. And set the path to training and test data in the shrec21.yaml file in the configs/ folder.

 train_data_dir: './path/to/Train-set/'
 test_data_dir: './path/to/Test-set/'

Training

• Online training:

bash ./scripts/train_online.sh

• Offline training:

bash ./scripts/train_offline.sh

Evaluation

• Online Evaluation:

bash ./scripts/eval.sh

The predicted results will be stored at ./experiments/shrec21/random_60/results/online_evaluation_results.txt. The evaluation scores can be calculated using the official Matlab code provided by the Shrec'21.

Some visual results:

• Example of reconstructed window:

• Example of predictions:

• Shrec'21 online confusion matrix:

• Shrec'21 offline confusion matrix:

We thank MAE and STGCN for making their code available.