A Study of Landmark-Augmented and Landmark-Free Approaches: Yoga Pose Recognition

In the realm of yoga pose recognition, this project delves into a comprehensive investigation of two distinct methodologies, namely landmark-augmented and landmark-free approaches, to enhance the accuracy and efficacy of pose recognition systems. The first approach leverages state-of-the-art pose estimation techniques employing MoveNet, MediaPipe, and YOLO V8 to precisely detect pivotal anatomical landmarks that define yoga postures. Subsequently, these detected landmarks are utilized as foundational cues for the classification process. On the other hand, the landmark-free approach directly employs cutting-edge deep learning architectures, specifically Vision Transformer, ResNet50, and VGG16, to classify yoga poses without the explicit utilization of detected landmarks. By juxtaposing these two methodologies, the project endeavors to uncover the comparative advantages, challenges, and limitations associated with each approach in the domain of yoga pose recognition. Beyond performance evaluation, this study aims to provide valuable insights into the most effective and robust techniques for accurately identifying and classifying yoga poses. The findings are poised to contribute to the development of sophisticated yoga pose recognition systems that hold promise in assisting yoga practitioners in achieving proper alignment and posture during their practice, ultimately promoting their overall well-being and practice effectiveness.