A real-time yoga pose classification system using MediaPipe pose estimation and PyTorch neural networks. This project achieves 95.4% validation accuracy in classifying yoga poses from images.

This project helps yoga practitioners get instant feedback on their poses without needing an instructor present. Built as a complete end-to-end ML pipeline, it demonstrates skills in computer vision, deep learning, and deployment.

As someone passionate about making yoga accessible, I built this to help practitioners verify their form in real-time. This project taught me:

• End-to-end ML engineering (data preprocessing → deployment)
• Production-ready computer vision pipelines
• Real-time pose estimation and classification
• Web application deployment with Streamlit

• High Accuracy: 95.4% validation accuracy on yoga pose classification
• Real-time Detection: Fast pose estimation using MediaPipe
• Image & Video Support: Works with both static images and video files
• Visual Feedback: Annotated pose visualization with confidence scores
• Interactive Web App: User-friendly Streamlit interface
• Production-Ready: Clean, modular code with proper train/test separation

The system consists of three main components:

1. Feature Extraction: MediaPipe extracts 33 3D landmarks (99 features total)
2. Neural Network: Fully connected network with dropout regularization
3. Classification: Multi-class softmax classifier with confidence scoring

Input (99 features: 33 landmarks × 3 coordinates)
    ↓
Linear(99 → 128) + ReLU + Dropout(0.3)
    ↓
Linear(128 → 64) + ReLU + Dropout(0.2)
    ↓
Linear(64 → num_classes)
    ↓
Softmax Output

Try the live app here!

Upload an image or video to see real-time pose detection and classification!

• Python 3.10+
• pip

# Clone the repository
git clone https://github.com/Pooja-Vachhad/AI-Yoga-Pose-Classifier.git
cd AI-Yoga-Pose-Classifier

# Install dependencies
pip install -r requirements.txt

This project uses the Yoga Pose Classification Dataset from Kaggle.

Dataset Details:

• 5 yoga poses: Downdog, Goddess, Plank, Tree, Warrior2
• High-quality images for training and validation
• Preprocessed using MediaPipe pose estimation

Dataset Structure:

dataset/
├── Downdog/
├── Goddess/
├── Plank/
├── Tree/
└── Warrior2/

To prepare the dataset:

1. Download from Kaggle
2. Extract to dataset/ folder
3. Run preprocessing script (see Usage below)

# Update the dataset path in creating_csv.py
# Then run:
python creating_csv.py

This extracts pose landmarks and creates pose_landmarks.csv.

python train.py

This will:

• Train the neural network for 50 epochs
• Display training progress with loss and accuracy
• Save yoga_pose_classifier.pth (model weights)
• Save label_encoder.pkl (class encoder)

# Update test_image_path in test.py
python test.py

streamlit run app.py

Then open http://localhost:8501 in your browser.

Web App Features:

• Upload images or videos
• Real-time pose detection with landmarks
• Confidence scores for predictions
• Side-by-side comparison
• Download processed videos

AI-Yoga-Pose-Classifier/
├── app.py                      # Streamlit web application
├── train.py                    # Model training script
├── test.py                     # Testing script for images
├── creating_csv.py             # Dataset preprocessing
├── requirements.txt            # Python dependencies
├── packages.txt                # System dependencies
├── yoga_pose_classifier.pth    # Trained model weights
├── label_encoder.pkl           # Class label encoder
├── README.md                   # Project documentation
└── dataset/                    # Training data (not included)
    ├── Downdog/
    ├── Goddess/
    ├── Plank/
    ├── Tree/
    └── Warrior2/

Solution:

• Added dropout layers (0.3 and 0.2) to prevent overfitting
• Implemented stratified train-test split for balanced data
• Result: 95.4% validation accuracy

Solution:

• Used MediaPipe's optimized pose estimation
• Lightweight neural network architecture (~50KB)
• Achieved real-time inference on CPU

Solution:

• Frame-by-frame pose detection
• Efficient video encoding with OpenCV
• Added download feature for processed videos

• Add more yoga poses (expand to 20+ poses)
• Real-time webcam support
• Mobile app deployment (Flutter/React Native)
• Pose correction suggestions
• Multi-person pose detection
• Export workout analytics