Yoga Pose Classification using CNN and MediaPipe

This repository contains an implementation of a deep learning approach for yoga pose classification using Convolutional Neural Networks (CNN) and MediaPipe for body keypoint detection.

Features

• Dataset: Utilizes a dataset of yoga poses with images processed using MediaPipe.
• Model Architecture: Implements a custom CNN model (YogaConvo2d) and compares it with transfer learning models like VGG16, InceptionV3, InceptionResNetV2, and NASNetMobile.
• Skeletonization: Uses MediaPipe to skeletonize images, improving classification accuracy.
• Training: Includes training scripts with loss and accuracy tracking.
• Evaluation: Provides detailed performance metrics, confusion matrices, and comparison with state-of-the-art models.

Dataset link: https://www.kaggle.com/datasets/niharika41298/yoga-poses-dataset

Setup and Usage

Clone the Repository

git clone [https://github.com/yourusername/yoga_pose_classification.git](https://github.com/shub-garg/Yoga-Pose-Skeletonization-and-Classification)
cd Yoga_Pose_Skeletonization_and_Classification

Launch Jupyter Notebook

jupyter notebook cnn-yoga.ipynb

Run All Cells

Execute all cells in the notebook to preprocess data, define models, and start the training process.

Results and Observations

Training and Validation Accuracy: Detailed accuracy and loss graphs for YogaConvo2d. Confusion Matrix: Displays the confusion matrix for model evaluation.

Results Table

Model	Accuracy (Validation)	Accuracy (Test)	Precision	Recall	F1 Score
YogaConvo2d (MediaPipe)	99.62%	97.09%	0.97	0.97	0.97
YogaConvo2d (Non-MediaPipe)	89.97%	89.36%	0.89	0.88	0.89
InceptionV3 (MediaPipe)	95.09%	94.39%	0.95	0.94	0.94
InceptionV3 (Non-MediaPipe)	94.98%	94.04%	0.94	0.94	0.94
NASNetMobile (MediaPipe)	96.23%	94.39%	0.94	0.95	0.95
NASNetMobile (Non-MediaPipe)	94.31%	92.77%	0.94	0.92	0.93
InceptionResNetV2 (MediaPipe)	96.60%	95.29%	0.95	0.95	0.95
InceptionResNetV2 (Non-MediaPipe)	89.30%	86.81%	0.87	0.86	0.86
VGG16 (MediaPipe)	97.36%	96.86%	0.97	0.97	0.97
VGG16 (Non-MediaPipe)	95.60%	95.53%	0.96	0.95	0.96

Example Plots Accuracy and Loss Curves

Confusion Matrix

For more details, please refer to the following paper:

@article{Garg:2023,
  Author = {Shubham Garg and Aman Saxena and Richa Gupta},
  Title = {Yoga pose classification: a CNN and MediaPipe inspired deep learning approach for real-world application},
  Year = {2023},
  Journal = {Journal of Ambient Intelligence and Humanized Computing},
  Volume = {14},
  Number = {12},
  Pages = {16551--16562},
  Publisher = {Springer},
  Howpublished = {\url{https://link.springer.com/article/10.1007/s12652-022-03910-0}}
} 

Link to paper.