Hierarchical pseudo-label learning for one-shot person re-identification

摘要

The aim of one-shot person re-identification (person re-ID) is to recognize the target person from abundant pedestrians with only one labeled image of each person in training samples. The major problems faced by one-shot person re-ID algorithms are information loss and label noise. In this work, we propose a novel approach targeting one-shot person re-identification, planning to solve both the problems. It overcomes the information loss problem by introducing a hierarchical pseudo-label assignment strategy to fully exploit the unlabeled data information. Further, we also enhance the independence of two network branches by multiplying them by different weights for co-training in our asymmetric mutual teaching framework, which can generate soft labels to reduce hard pseudo-label noise. Extensive experiments on three benchmark datasets suggest that our proposed method significantly outperforms the existing state-of-the-art algorithm of mAP, respectively on Market-1501 and on DukeMTMC-reID, by 31.5 and 18.1. Our method can also be applied to the few-shots setting and achieve respectable performance.