Review of an Evolved DNN Architecture EfficientNet for Yoga Pose Detection Problem

Yoga is a popular form of exercise that has been rehearsed for centuries. For many sports or exercises, such as yoga, golf, rugby, table tennis, etc., several automatic or semi-automatic training systems have been developed. Perform yoga with correct postures is the critical challenge that individuals find while practicing them. It is essential to practice yoga activities correctly, otherwise, it may have enormous affects. Recognition of posture is a challenging task because of the unavailability of the proper dataset to be used as reference. In this paper, we have presented a yoga pose detection system using computer vision and deep learning techniques. This manuscript uses a transfer learning approach with a highly performing EfficientNet architecture. The network has been trained for eight different models named B0 to B7. The architectural differences and the model performance have been demonstrated for these 8 models on a classification problem of yoga pose detection. EfficientNet B7 deliver the best accuracy of 97.61% among all 8 models. We envision that this system could be used as a tool to assist yoga practitioners in monitoring and improving their form and technique.