Yoga Pose Recognition through Fine-Tuned ResNet50V2 Model with Projection Head
Shanvi Chauhan, Priyanka Sumit Chetwani, Shubham Mahajan · 2024
Yoga, an ancient practice known for its advantages on physical and mental well-being, consists in a vast range of poses that need exact classification for use in automated fitness systems and research. This work introduces a fresh method based on advanced deep-learning techniques to categorize certain yoga positions. We used a ResNet50V2 model connected with a projection head for feature extraction combined with a supervised contrastive loss function to improve the discriminative capability of the model. Comprising 1551 total images, the study comprises training (666 photos), validation (215 images), and testing (470 images) phases. By properly differentiating between different poses, the supervised contrastive learning method greatly enhanced the performance of the model. With an outstanding 97% classification accuracy, our approach proved capable for precise recognition of yoga poses. This development promotes individualized yoga practice and may help to prevent injuries, thus improving health and well-being. It also fits the aim of excellent education since it provides a means for fitness facilities and educational institutions to improve their yoga training courses. By using innovative deep learning methods and using them in easily available fitness solutions, the research promotes creativity. This progress not only helps to create automated yoga training systems but also creates opportunities for more study in the field of position classification with deep learning approaches.