Lightweight Human Pose Estimation Algorithm Based On Structure Reparameterization And Knowledge Distillatio

Human pose estimation is an important foundational task in the computer vision field.The goal is to locate the coordinates of key points on the human body in an input image.Currently,high-performance human pose estimation models have low inference efficiency,making it difficult to implement the algorithms.Therefore,the research focuses on the lightweighting of2 D human pose estimation algorithms.The work can be divided into the following two parts:(1)A lightweight human pose estimation model based on structural reparameterization is proposed.Currently,mainstream lightweighting methods often using depth wise separable convolutional modules to reducing the parameter and computational complexity.However,when experimenting with such models on GPU devices,it was found that the high memory access of depth wise separable convolutions limits the inference speed.This paper uses 3x3 convolutions to construct a lightweight network and optimizes the learning process of 3x3 convolutions using structural reparameterization.To improve the efficiency of multi-scale feature fusion,a lightweight multi-scale feature fusion module is designed.Experiments show that the proposed method improve the inference speed by an average of 20% compared to the current advanced lightweight model Lite HRNet,achieving 1.6% improvement in PCKh@0.5on the MPII dataset and 1.0% improvement in AP on the COCO dataset.(2)Based on the lightweight model Rep HRNet,knowledge distillation technology is introduced.To improve the efficiency of two stage knowledge distillation framework,an online knowledge self-distillation framework is designed,using a multi-branch high-resolution network as the teacher model,fusing multi-branch output heatmaps as soft labels,and finally adopting one of the branches as the student model.The learning between students should also be considered,low-level knowledge between students should also be considered in contrast to just using the high-level knowledge of soft labels.In the proposed method,we integrate teacher and students’ learning into one single framework.The student models’ low-level features supervise each other in the form of attention information.The method in this section achieving a 1.0% improvement in PCKh@0.5 on the MPII dataset and 1.3% improvement in AP on the COCO dataset without increasing the inference cost.