e-lab / Actions

Gesture Recognition from Videos

This repository contains scripts to train deep neural network for gesture recognition using video dataset. Our aim is to use the temporal coherency present in video data and stabilize neural network's output probabilities.

Dependencies

• matplotlib: used by scikit-video

pip install matplotlib

• scikit-video: accessing images / videos

pip install sk-video
pip install scikit-image --upgrade

• tqdm: progress bar

pip install tqdm

conda config --add channels conda-forge
conda update --all
conda install tqdm

Input data structure

data/
├── class_1
│   ├── video_1.mp4
|   ...
|   └── video_m.mp4
└── class_2
    ├── video_1.mp4
    ...
    └── video_n.mp4

Step 1: Prepare dataset

Folder containing videos (following the above structure) will be used as an input for prepData.py script. This script will convert these videos into tensors with .pyt extensions.

Step 2: Train neural network

Use main.py script to train your neural network. Make sure the input image resolution is the same as that provided in the above step. Also, dataset location, place to save model, batch size are some of the inputs which must be provided to the training script.

python3 main.py --data /data/in/tensor --save /save/trained/model