Learn from Teachers: A Teacher Learner Yoga Pose Classification and Scoring Network

Abstract:
Nowadays, computer vision is more and more widely applied in fitness and healthcare. However, little research has been done on yoga in this area, even though Yoga is a very popular sport. To fill this gap, this paper proposes a teacher-learner convolutional network, namely TL-CNN model, to realize yoga pose classification and scoring. The TL-CNN model is a convolutional neural network designed based on the Siamese network. It consists of two neural networks with shared weights, namely Teacher-CNN and Leaner-CNN. Input an image pair, by calculating the cosine similarity of two embedded vector output of the Teacher-CNN and the Learner-CNN, the similarity score can be obtained. At the same time, the yoga pose classification result of the image inputting in Learner-CNN can be also obtained. In this paper, four classic yoga pose categories are selected as datasets. Satisfactory test results have been achieved in both Yoga pose scoring and classification tasks. More specifically, the classification accuracy rate reaches 1, the similarity score of the same class of pose pairs is greater than 0.99, and the similarity score of different class pose pairs is less than -0.2.