An Unsupervised Person Re-Identification Method Based on Intra-/Inter-Camera Merger

In criminal investigation, intelligent monitoring, image retrieval and other fields, person re-identification has always been a hot and significant research topic. Since most of the existing methods rely on labeled datasets, the lack of labels makes unsupervised person re-identification technology more challenging. In order to overcome the problem of the lack of labels, a new framework to generate reliable pseudo labels is proposed as supervision information for existing supervised person re-identification models. Assuming that images taken by the same camera mainly vary in pedestrian’s physical appearances rather than backgrounds, then images taken by different cameras vary in backgrounds. To eliminate effects brought by the differences in image background, images are divided into several domains by camera Serial numbers as the first step. Then we construct undirected graphs for each camera with Euclidean distance of image pairs, and there will be an edge only when two images are close enough. The images in one maximum analogous set are regarded as the same person. Then we merge maximum cliques from different cameras simply with their cosine distance and assign pseudo labels. The proposed framework runs in an unsupervised manner, and the proposed method can obtain higher accuracy than the other related methods on Market1501 and DukeMTMC-ReID datasets, which further shows the effectiveness of the proposed method.