HierarchicalNets for multi level hierarchical classification of yoga poses

Human pose estimation has long been an actively studied subject in computer vision (CV), with substantial applications in healthcare. It assists in evaluation and monitoring, enabling personalized rehabilitation programs and injury prevention. When applied to yoga, human pose estimation provides accurate alignment guidance throughout practice, reducing the risk of injury and promoting correct posture. This combination enhances physical therapy outcomes, supports overall wellness, and encourages the maintenance of a healthy and efficient yoga practice for general health improvement. We present a unique approach that combines cutting-edge vision transformer technologies with fine-grained hierarchical pose classification. Specifically, we introduce four hierarchical vision transformer architectures designed to incorporate hierarchical class label information into classification models. The proposed method aims to produce more precise classification results by identifying subtle differences between similar poses, thereby improving classification accuracy for yoga postures. The current state-of-the-art (SoTA) Top-1 accuracy for Level 1, 2, and 3 classification tasks is 89.81%, 85.10%, and 79.35%, respectively. Our proposed models surpass these benchmarks, achieving Top-1 accuracy scores of 96.79%, 95.64%, and 93.07%, respectively.