Pedestrian Attribute Recognition Model based on Adaptive Weight and Depthwise Separable Convolutions

Abstract

Pedestrian attribute recognition is a technology that uses computer vision technology to determine some observable external characteristics of pedestrians in an image or video sequence, such as gender, age, clothing, and equipment, etc. It is widely used in many fields, such as intelligent security, clustering based passenger flow statistics in public places, human-computer interaction, album clustering and so on, It has high research value in the field of computer vision. Due to the complexity and diversity of pedestrian attributes, the convergence rate of each attribute is not consistent in the process of model training. Therefore, the manual training with fixed weight set by experience cannot adapt to the diversification of attributes, resulting in poor performance of the model. In recent years, in some researches, a multi-branch model is usually needed to solve this problem through multiple trainings. This method has a lot of parameter redundancy, which affects the efficiency of the model. The aim of this paper is to improve the accuracy by using a dynamic adaptive weight loss model without multiple branches in a single training. The depthwise separable convolutions is used to further improve the model recognition efficiency. After training in the PETA dataset and adding dynamic adaptive weight loss, the accuracy of the model is similar to that of some multi-branch models, and the recognition efficiency is greatly improved.