Pedestrian re-identification method based on two-way mutual promotion deentanglement learning

Abstract

The invention discloses a pedestrian re-identification method based on two-way mutual promotion deentanglement learning, and belongs to the field of computer vision. The method comprises the steps of obtaining a content encoder with extraction domain invariant features through a training process, and performing re-identification on pedestrians in a target domain test sample by using the content encoder in a test process. Compared with a traditional pedestrian re-identification method, the method of the invention is simple and effective and has higher practical value. And the invention shows more excellent performance on different data sets.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a pedestrian re-identification method based on two-way mutual promotion disentanglement learning. Background technique [0002] Pedestrian re-identification is a technology to judge whether the pedestrian images captured by cross-view cameras are the same pedestrian. This technology is an important part of intelligent monitoring. It has important applications in tracking criminal suspects and finding missing persons. It has attracted the attention of many researchers. With the rapid development of deep learning, person re-identification has made significant research progress in recent years and achieved excellent recognition performance. However, these performances are generally obtained by supervised training on the same dataset and testing on that dataset. If these methods are directly deployed to real-world scenarios, the recognition performance will suffer a sharp drop due to...

AI Technical Summary

Problems solved by technology

[0005] 1. The training method adopts supervised training. When it is directly deployed to the real scene, the recognition performance will drop sharply due to the domain shift between the training data and the test data.

[0006] 2. Based on the method of pseudo-label prediction, each target sample that needs to participate in training has a positive sample, which is inconsistent with the actual application scenario. If it is directly deployed in an actual scene with very few positive sample pairs, the performance will drop sharply

[0007] 3. The method based on additional model assistance requires the assistance of additional models, which greatly affects the recognition efficiency of the re-ID model

[0008] 4. Traditional domain-invariant feature extraction methods have relatively low recognition performance on public datasets

Images