Research on Lightweight Human Pose Estimation Model Based on Knowledge Distillation

Abstract

Most of the existing state-of-the-art human pose estimation methods mainly pursue improving the precision of human pose estimation, ignoring the model's complexity and making it impossible to deploy it on resource-constrained devices. In addition, some other methods overly pursue the model's lightweight, which significantly decreases the model's precision and affects the experience of practical applications. However, in reality, fast and accurate applications are often more popular, so this area of research is of some practical significance. To achieve lightweight deployment and better practical application, A lightweight human pose estimation model based on knowledge distillation is proposed, which firstly uses group convolution and channel shuffle operation to lighten the HRNet model to obtain the S-HRNet model. Secondly, a response-based knowledge distillation algorithm is proposed to use the pre-trained HRNet as the teacher model and S-HRNet as the student model with the help of an offline knowledge distillation strategy, and the prediction results of the teacher model are used to supervise the learning of the student model together with the real labels. Finally, this paper conducts a large number of experiments on the MS COCO dataset proposed by Microsoft, and the experimental results show that the model in this paper has a higher average precision and lower number of parameters and computations compared to the other models mentioned.