Learning discriminative region representation for person retrieval

Abstract

Region-level representation learning plays a key role in providing discriminative information for person retrieval. Current methods rely on heuristically coarse-grained region strips or directly borrow pixel-level annotations from pretrained human parsing models for region representation learning. How to learn a discriminative region representation within fine-grained segments while avoiding expensive pixel-level annotations is rarely discussed. To that end, we introduce a novel identity-guided human region segmentation (HRS) method for person retrieval. Via learning a set of distinct region bases that are consistent across a given dataset, HRS can predict informative region segments by grouping intermediate feature vectors based on their similarity to these bases. The predicted segments are iteratively refined for discriminative region representation learning. HRS enjoys two advantages: (1) HRS learns region segmentation using only identity labels, making it a much more practical solution to person retrieval. (2) By jointly learning global appearance and local granularity cues, HRS enables a comprehensive feature representation learning. We verify the effectiveness of the proposed HRS on four challenging benchmark datasets of Market1501, DukeMTMC-reID, CUHK03, and Occluded-DukeMTMC. Extensive experiments demonstrate superior performance over the state-of-the-art region-based methods. For instance, on the CUHK03-labeled dataset, the performance increases from 74.1% mAP and 76.5% rank-1 accuracy to 81.5% ( +7 . 4 %) mAP and 83.2% ( +6 . 7 %) rank-1 accuracy.