Cross Domain Person ReID Based on Symmetric Coding and Pedestrian Similarity

Abstract

AbstractPerson re-identification (ReID) aims to find images of the same pedestrian in cross camera scenarios. Recent years have witnessed extensive studies on person ReID by exploiting machine learning techniques under single domain settings. In this paper, we propose a novel cross domain person ReID method based on symmetric coding and pedestrian similarity. Specifically, we first introduce a symmetric coding technique for pedestrian attributes, and compute the attribute similarity matrix. Then, we extract pedestrian features by utilizing convolutional neural networks, and compute the feature similarity matrix. By fusing the pedestrian attribute similarity and pedestrian feature similarity, we obtain the fusion similarity accordingly. The attribute-tensor graph can be represented by the edges composed of the pedestrian similarity, which is produced by pedestrian attribute similarity and pedestrian feature similarity. Finally, the person ReID results can be obtained by similarity ranking. We conduct comprehensive experiments to evaluate the performance of our proposed algorithm for cross domain person ReID tasks, in which the encouraging results validate the efficacy of our proposed cross domain person ReID method.KeywordsCross domainSymmetric codingPedestrian similarityPerson ReID