Person Re-Identification Method Based on Joint Fusion of Saliency Multi-Scale Features

Abstract

The existing person re-identification methods are limited by multiple factors, such as complex background information and occlusion, which reduces the discrimination and robustness of extracted features, leading to a low re-identification accuracy.To address the problem, this paper proposes a new method called SMC-ReID based on saliency detection and collaborative fusion of multi-scale features.The method employs saliency detection to extract discriminative feature areas in pedestrians, and the saliency features are fused with global features.Then the features are cut at different scales, and collaboratively fused to ensure the continuity of the cut features.Finally, the three loss functions are combined to learn based on the differences between global and local features.In the inference stage, the features of each scale are reduced to the same dimension, and fused into new feature vectors for similarity measurement. Experimental results on the public datasets for person re-identification, such as Market1501, DukeMTMC-reID and CUHK03, show that the features extracted by the proposed method have strong distinguishability and robustness, and the method has higher identification accuracy than SVDNet, PSE+ECN and other advanced algorithms.