Part-Weighted Deep Representation Learning for Person Re-Identification

Person Re-identification (ReID) is to identify the same person across different cameras. It is challenging due to various appearance presented by person images, especially when some person body parts are missing caused by detectors or occlusions. To overcome these challenges, in this paper we propose a Part-Weighted Deep (PWD) model. Our deep multi branch network jointly learns complementary discriminative features from aligned whole body image and body part regions, which are acquired by a pose estimation algorithm. To estimate the reliability of features extracted from local body parts, a Feature Fusion sub-Net (FFN) is designed to learn the weight of each body part, which is used to fuse representation in training stage and aggregate part similarity scores between two person images in testing stage. Extensive experiments demonstrate the significant performance improvements of our PWD model over the state-of-the-art methods on two benchmarks (Market-1501, Duke).