This repository provides some unofficial pytorch implementations of awesome works for surface electromyography (sEMG) based hand gesture recognition.

Please star this repo if you find our work is helpful for you.

• Network architectures to be reproduced
  • Multi-stream CNN (WeiNet)
  • XceptionTime
  • BiTCN

The code is developed using python 3.7 on Ubuntu 20.04. NVIDIA GPU is needed.

The experiment are taken on the Ninapro dataset. The first sub-dataset DB1 and second sub-dataset DB2 are ultilized.

1. Firstly download the Ninapro DB1 and Ninapro DB2 datasets. And then extract data files from the zip files, we provide two jupyter notebooks extractFile_db1 / extractFile_db2 for extracting DB1 / DB2 respectively. Your directory tree should look like this:

${ROOT}/data/ninapro
├── db1
|   |—— s1
|   |—— s2
|   |   ...
|   └── s27
|       |—— S27_A1_E1.mat
|       |—— S27_A1_E2.mat
|       └── S27_A1_E3.mat
└── db2
    |—— DB2_s1
    |—— DB2_s2
    |   ...
    └── DB2_s40
        |—— S40_E1_A1.mat
        |—— S40_E2_A1.mat
        └── S40_E3_A1.mat

2. We provide two jupyter notebook scripts process_db1 / process_db2 for convert the mat files to txt files. After convertion, your directory tree should look like this:

${ROOT}/data/ninapro
├── db1_processed
|   |—— s1
|   |—— s2
|   |   ...
|   └── s27
|       |—— emg.txt
|       |—— rerepetition.txt
|       └── restimulus.txt
└── db2_processed
    |—— DB2_s1
    |—— DB2_s2
    |   ...
    └── DB2_s40
        |—— emg.txt
        |—— rerepetition.txt
        └── restimulus.txt

1. Clone this repo
2. Install dependencies:

   pip install -r requirements.txt
   

• Train network multi-stream CNN on Ninapro DB1 dataset:

  sh scripts/train_db1_MSCNN.sh
  

• validate net

The comparison between reported accurary in paper and reprodecud accuracy are demonstrated as Table.1.

If you have any questions, feel free to contact me through [email protected] or Github issues.