REGION-PARTITION BASED BILINEAR FUSION NETWORK FOR PERSON RE-IDENTIFICATION

Abstract

Person Re-Identification (ReID) aims to match people across disjoint camera views. Feature representation and matching are two critical components in person ReID task. In this paper, we introduce a region-partition based bilinear network (RPBi-Net), aiming to capture both global and local information simultaneously. Firstly, a novel Part Box Estimation (PBE) sub-network is embedded to operate region partition on original image. Considering the different importance of human parts, we propose a weighted region partition loss when learning PBE. Secondly, a two stream convolutional neural network is built to learn high-level feature representation from both the whole and partitioned human body. Finally, the learned local and global features are fused in a compact bilinear way, so as to acquire a final descriptor for matching pedestrians. Experimental validation on three benchmark datasets, i.e., CUHK01, CUHK03, Market1501, demonstrates that our model significantly outperforms the state-of-the-art methods.