Learning Incremental Triplet Margin for Person Re-Identification

• DOI: 10.1609/aaai.v33i01.33019243
• Publication Date: 2019-08-19T07:32:09Z
• Authors: Yingying Zhang; Qiaoyong Zhong; Liang Ma; Di Xie; Shiliang Pu
• Keywords: Margin (machine learning); Discriminative model; Feature (linguistics); Identification; Sample (material); Similarity (geometry)

Person re-identification (ReID) aims to match people across multiple non-overlapping video cameras deployed at different locations. To address this challenging problem, many metric learning approaches have been proposed, among which triplet loss is one of the state-of-the-arts. In work, we explore margin between positive and negative pairs triplets prove that large beneficial. particular, propose a novel multi-stage training strategy learns incremental improves effectively. Multiple levels feature maps are exploited make learned features more discriminative. Besides, introduce global hard identity searching method sample identities when generating batch. Extensive experiments on Market-1501, CUHK03, DukeMTMCreID show our approach yields performance boost outperforms most existing state-of-the-art methods.