YogaPose-8 : A Diverse Yoga Pose Dataset for Classification and Analysis

Proper identification of yoga-poses is important in ensuring safe and effective practice, yet most existing machine learning techniques often raise several data privacy concerns and face challenges in acquiring diverse datasets, particularly concerning data availability. Our study is the first to employ Flower Federated Learning for classifying eight distinct yoga-poses in a way that resolves these issues, by allowing multiple clients to train a shared model collaboratively without revealing personal information. This method leverages Convolutional Neural Networks (CNNs) for feature extraction and combines knowledge from clients’ private datasets while ensuring confidentiality and improving data availability. We utilize a carefully curated dataset of 7442 images of eight poses, collected from users and various websites, to train our model. Our comprehensive training achieves an impressive accuracy of 99.2%, demonstrating the efficacy of federated learning. Our findings show that federated learning can preserve data privacy while delivering performance levels comparable to centralized models. Additionally, the federated architecture provides robustness against changes in data distribution across client devices. This research offers a novel approach to yoga-pose identification, paving the way for precise techniques that address both data privacy and availability concerns, and acting as a foundational step towards effective yoga-pose correction.