This repository contains the source code for our paper: Spatio-Temporal Sparse Graph Convolution Network for Hand Gesture Recognition
Our approach
The main steps include:
-
Decoupled learning of directed and sparse spatial and temporal interactions, using attention mechanism and asymmetric convolutions.
-
Acquiring cross-spatial-temporal graph representations, by capturing both spatial-temporal and temporal-spatial interactions.
- ...
Create and activate conda environment:
conda create -n stsgcn_env python=3.10
conda activate stsgcn_env
Install all dependencies:
pip install -r requirements.txt
By default, the 3D hand joints training dataset is structured as follows for each sequence (see example in ./datasets/
):
## frames 1
x1, y1, z1
x2, y2, z2
...
...
x21, y21, z21
## frame 2
x1, y1, z1
x2, y2, z2
...
...
x21, y21, z21
## frame t
...
## frame T
x1, y1, z1
x2, y2, z2
...
...
x21, y21, z21
- Download the SHREC’17 Track Dataset , the Briareo Dataset , the SHREC'21 Dataset and the Extracted landmarks IPN Dataset
- Set the path to you dataset in
train.py (line 107)
- Run the following command to train the model
python train.py --params
If you find this repo useful, please consider citing our paper
ref
This code borrows from SGCN.