Knowledge Compensation Network with Divisible Feature Learning for Unsupervised Domain Adaptive Person Re-identification

Abstract

Due to the large domain shift and the discriminative feature learning with unlabeled datasets, unsupervised domain adaptation (UDA) for person re-identification (re-ID) still remains a challenging task. Some current methods adopt a clustering-based strategy to assign pseudo labels to the unlabeled samples in target domains for classification. However, the rich knowledge of the model in different training stages is not fully utilized in those methods and the pseudo labels generated by clustering algorithms inevitably contain noise, which would limit the performance of re-ID models. To tackle this problem, a Knowledge Compensation Network with Divisible feature learning (KCND) is proposed in this paper, which aggregates the past-to-present knowledge of models from training samples for discriminative feature learning and resists the label noise produced by clustering. Also, a novel compensation-guided softened loss is developed to enhance the generalization and robustness of re-ID models. Our experimental results on large-scale datasets (Market-1501, DukeMTMC-reID and MSMT17) have demonstrated the performance of KCND is better than other methods in terms of the mAP and CMC accuracy.