Develop a real-time dynamic hand gesture recognition system using the HaGRID and JESTER datasets. It includes detection and classification components for accurate performance. Below are the project structure, setup, and execution details.
- setup.sh : bash file to setup environment and download dataset
- detect_train.py : train detection models
- train.sh : bash file to train classify models from scratch or pre-trained models
- detect_eval.py : evaluate detection models
- classify_eval.py : evalutate classification models
- program.py : run the program on real-time camera
In this project, we use HaGRID dataset for detection task. You can download the version of the dataset for YOLO here:
For classification task, we use JESTER dataset. You can download it here:
Setup directories: Hand_Gesture/src.
After clone this repository in src, run this command to install requirement packages, download and setup dataset:
./setup.sh
CAUTION: You must have at least 70GB of memory available.
To train, evaluate models as well as run the program, you have to manually tune parameters from the source code.
Train:
python detect_train.py
Evaluate (image or video mode):
python detect_eval.py
Run this command to train model as well as save metrics:
You can set small_version = True in classify_train.py to run a demo with a small version of the dataset.
./train.sh
Evaluate:
python classify_eval.py
Run the model with real-time camera:
python program.py
Quoc Khanh