yishou
one-hand gesture recognition
2 min readupdated
This compact MLP (multilayer perceptron) model identifies Chinese digital hand gestures from 1 to 10 using 21 key landmarks extracted via MediaPipe. It takes 63 normalized features (x, y, z coordinates) as input and outputs one of 11 classes (digits 1 to 10 or no gesture). This solution is ideal for developers building gesture interfaces or interactive applications controlled via a webcam.
Key features
- 11 classes: recognizes numbers from 1 to 10 using single fingers, open palms, fists, and other configurations, plus a class for other gestures.
- 96.4% accuracy: validated on completely holdout recordings (4,051 validation and 7,052 training frames).
- Lightweight: simple MLP architecture runs quickly in real time alongside the MediaPipe library.
What you can do
- Interactive demos: launch the built-in command-line demo to test the model in real time using a webcam.
- Gesture control: integrate single-hand number recognition into Python scripts for touchless application control.
- MediaPipe experiments: use the model as a ready-made foundation for computer vision projects focused on hand pose analysis.
Example
Use the following command from the repository to install the package and run the built-in gesture recognition demo:
pip install git+[https://github.com/sparkup/yishou](https://github.com/sparkup/yishou)
yishou demo --model dworsleytonks/yishouRead more
DetailsValidation accuracy: 96.4% (based on 4,051 validation and 7,052 training frames).
Limitations: all data was recorded by a single person using one webcam, so accuracy may drop with different hands, cameras, or lighting conditions.










Comments