Multi-stage Transfer Learning Based Yoga Pose Recognition Using CNN

In this paper, we propose a novel yoga pose recognition technique utilizing the frame work of multi-stage transfer learning. We use bottom up approach to obtain heatmaps and part affinity fields for each joint location of a person. These serve as cues for training of subsequent stages. We propose a modification of the well known computationally efficient PeleeNet architecture. An additional branch called “2D branch" which serves the purpose of obtaining heatmaps and part affinity fields, has been used in the initial phase of the training. The objective of this study, to recognize the yoga pose being performed, is carried out at three different hierarchies using hierarchical classification. In this study, we report results on 17,525 images from the publicly available yoga dataset, “Yoga-82". We obtained yoga pose Top-1 classification accuracy of 90.11% over 6 pose classes (Yoga-6), 87.57% over 20 pose classes (Yoga-20) and 83.33% over 82 pose classes (Yoga-82). The number of parameters required for the computation is approximately less by a factor of 30% while resulting in a peak improvement of 4% as compared to the best reported state of the art methodologies.