Human semantic parsing for person re-identification

Abstract

Person re-identification is a challenging task mainly dueto factors such as background clutter, pose, illuminationand camera point of view variations. These elements hinder the process of extracting robust and discriminative representations, hence preventing different identities from being successfully distinguished. To improve the representation learning, usually local features from human body partsare extracted. However, the common practice for such aprocess has been based on bounding box part detection.In this paper, we propose to adopt human semantic parsing which, due to its pixel-level accuracy and capabilityof modeling arbitrary contours, is naturally a better alternative. Our proposed SPReID integrates human semanticparsing in person re-identification and not only considerably outperforms its counter baseline, but achieves stateof-the-art performance. We also show that, by employinga simple yet effective training strategy, standard populardeep convolutional architectures such as Inception-V3 andResNet-152, with no modification, while operating solelyon full image, can dramatically outperform current stateof-the-art. Our proposed methods improve state-of-the-artperson re-identification on: Market-1501 [48] by ~17% inmAP and ~6% in rank-1, CUHK03 [24] by ~4% in rank-1and DukeMTMC-reID [50] by ~24% in mAP and ~10% inrank-1.