Yoga Pose Live Detection

Real-time yoga pose classification in the browser. The model runs client-side with TensorFlow.js — the webcam feed never leaves the machine.

What's in it

Path	Role
index.html	The live detection page
learn.js	Loads the model and runs inference on the webcam stream
modelCreation/dataCollection.js	Captures training samples from the webcam
modelCreation/modelTraining.js	Trains the classifier in-browser
modelCreation/modelDeploy.js, modelDeploy2.js	Exports the trained model
modelv2/model2.json + model.weights2.bin	The trained model, committed and ready to load

The whole loop — collect, train, deploy, run — happens in the browser, so there is no Python step anywhere.

Tech stack

• TensorFlow.js for training and inference
• Webcam capture via getUserMedia
• Plain HTML, CSS and JavaScript

Running it

The page loads model files over fetch, so opening index.html directly will fail on CORS. Serve the folder instead:

python -m http.server 8000

Then open http://localhost:8000. Allow camera access when prompted.

Any static server works, for example:

npx serve

Training your own

1. Open the data collection page and record samples for each pose
2. Run training — it happens live in the browser
3. Export, and drop the output into modelv2/

Known gaps

• Accuracy depends heavily on lighting and camera angle.
• The pose set is fixed to whatever the committed model was trained on.

License

See LICENSE in this repository.