An Improved Approach for Yoga Pose Estimation of Images
G. Shirisha, Neha R Bhat, Akshata S Hamasagar, Ananya A Hosamani, Priyadarshini C. Patil · 2024
This project focuses on advancing yoga pose estimation using deep learning. The dataset undergoes thorough preprocessing to ensure integrity, addressing issues like corrupted images. Drawing upon a pretrained VGG16 network and custom layers, the model is tailored for yoga pose intricacies. The model is evaluated on both training exercises and real-world tests, guiding further enhancements. The trained model is saved for future use, and testing involves systematic validation on new images using a translation dictionary for result interpretation. Our proposed model has achieved the accuracy of 97.30% with a validation loss of 52.82% and validation accuracy of 98.82% for Yoga Set 1 and accuracy of 97.48% with a validation loss of 55.63% and validation accuracy of 97.72% for Yoga Set 2. We expanded pose prediction from three to six poses using models. The high accuracy achieved suggests potential applications in not only pose recognition but also in providing valuable insights for improving form and technique. The project's outcomes contribute to computer vision and pave the way for applications in realtime feedback systems, promising advancements in yoga practice enhancement.