Multiple Uses of Global and Local Features for Person Re-identification

Abstract

Person re-identification has been extensively studied in recent years and has made great progress. Many papers propose a lot of effective methods to improve the accuracy of the person re-identification. However, there are still many problems that remain unsolved. For example, persons are often occluded by obstacles or other persons, leading to loss of the complete person information, and changes in person behaviors or postures make it difficult to identify. In this paper, we propose a person re-identification algorithm that repeatedly uses global feature information and local feature information for mutual supervised learning. The algorithm consists of two parts, person alignment branch and spatial channel feature branch. First, for person alignment branch, we use global feature information and local feature information to correct misaligned person pictures, and calculate the shortest distance to match the right part of the images. For the spatial channel feature branch, the spatial features are segmented to obtain the local feature information of the person image. At the same time, the spatial feature information is extended using the convolution layer and divided to obtain global feature information of the person image. The global feature information and local feature information are used to calculate the spatial channel feature loss. So that the network can learn better discriminative features through the global information and local information repeatedly. The experimental results show that, on the market-1501 and Duke datasets, the algorithm in this paper obtains good experimental results, has strong robustness, and has greatly improved the rate compared with the existing person re-identification algorithms.