Unsupervised person re-identification via re-ranking enhanced sample-specific metric learning
Heng Zhao, Zhenjun Han, Zhaoju Li, Fei Qin · 2017
Despite of the great progress of image-based person re-identification, most existing methods use supervised metric learning to build re-identification models and thus require repeated human effort to annotate sample pairs from non-overlapping cameras. In this paper, we propose an unsupervised sample-specific metric learning approach (SSML) to alleviate this problem. Specifically, using samples those are negatives (with a high probability) to the query samples, we train a local metric for each query sample following the max-margin learning theory. Moreover, a KNN intersection re-ranking (KIRR) method is used to further decrease the ambiguity of samples and aggregate the re-identification performance. With experiments on three widely used person re-identification datasets: VIPeR, CUHK01, and PRID, we demonstrate that the proposed approach is simple but effective.