Self-similarity Grouping: A Simple Unsupervised Cross Domain Adaptation Approach for Person Re-identification

• DOI: 10.48550/arxiv.1811.10144
• Publication Date: 2018-01-01
• Authors: Yang Fu; Yunchao Wei; Guanshuo Wang; Yuqian Zhou; Humphrey Shi; Thomas S. Huang
• Keywords: Similarity (geometry); Identification; Code (set theory); Domain Adaptation

Domain adaptation in person re-identification (re-ID) has always been a challenging task. In this work, we explore how to harness the natural similar characteristics existing samples from target domain for learning conduct re-ID an unsupervised manner. Concretely, propose Self-similarity Grouping (SSG) approach, which exploits potential similarity (from global body local parts) of unlabeled automatically build multiple clusters different views. These independent are then assigned with labels, serve as pseudo identities supervise training process. We repeatedly and alternatively such grouping process until model is stable. Despite apparent simplify, our SSG outperforms state-of-the-arts by more than 4.6% (DukeMTMC Market1501) 4.4% (Market1501 DukeMTMC) mAP, respectively. Upon SSG, further introduce clustering-guided semisupervised approach named ++ one-shot adaption open set setting (i.e. number unknown). Without spending much effort on labeling, can promote mAP upon 10.7% 6.9%, Our Code available at: https://github.com/OasisYang/SSG .