Real-time hand gesture recognition deployed with TensorRT.
All commands are ran from Thanos-Project root
pip install -r requirements.txt
- Add this directory in your PYTHONPATH
- Download IPN Hand dataset
- Create
dataset_config.json
in theThanos-Project
directory with key-value:- "ipn": "path/to/ipn/root"
- Download annotation
ipnall.json
from official repository, put it inpath_to_ipn_root/annotations/
A Transformer-Based Network for Dynamic Hand Gesture Recognition
python thanos\trainers\train_on_ipn.py CONFIG_PATH
Example
python thanos\trainers\train_on_ipn.py thanos\trainers\expe\default_config.py
Configuration: thanos\trainers\expe\default_config.py
- Input image size: 240x240
- Sequence length: 22 frames
- Temporal stride: 2
- Backbone: Resnet18
- Batch size: 4
- Accumulated gradient batch: 4
- Learning rate: 1e-4
- Number of epochs: 20
- Validation accuracy: 0.75
- Model predicts well some gesture.
- Model can distint non-gesture movement vs hand gesture (for example: drinking water).
- There are some early-detections which lead to a reduce in accuracy.
- Improve accuracy for isolated gesture
- Add hand segmentation
- Avoid early-detection
- Improve latency in inference by preprocessing image on GPU