Hierarchical Bi-Directional Feature Perception Network for Person Re-Identification

Abstract

Previous Person Re-Identification (Re-ID) models aim to focus on the most\ndiscriminative region of an image, while its performance may be compromised\nwhen that region is missing caused by camera viewpoint changes or occlusion. To\nsolve this issue, we propose a novel model named Hierarchical Bi-directional\nFeature Perception Network (HBFP-Net) to correlate multi-level information and\nreinforce each other. First, the correlation maps of cross-level feature-pairs\nare modeled via low-rank bilinear pooling. Then, based on the correlation maps,\nBi-directional Feature Perception (BFP) module is employed to enrich the\nattention regions of high-level feature, and to learn abstract and specific\ninformation in low-level feature. And then, we propose a novel end-to-end\nhierarchical network which integrates multi-level augmented features and inputs\nthe augmented low- and middle-level features to following layers to retrain a\nnew powerful network. What's more, we propose a novel trainable generalized\npooling, which can dynamically select any value of all locations in feature\nmaps to be activated. Extensive experiments implemented on the mainstream\nevaluation datasets including Market-1501, CUHK03 and DukeMTMC-ReID show that\nour method outperforms the recent SOTA Re-ID models.\n