Unsupervised Domain Adaptation via Attention Augmented Mutual Networks for Person Re-identification

Abstract

Supervised learning has limited generalization ability across scenes due to its high cost of data annotation, and unsupervised learning and unsupervised domain adaptation have become the hot topics in recent years. With the applications of deep learning in the field of unsupervised domain adaptation (UDA) for person re-identification, pseudo label methods via clustering techniques have become the mainstream route. However, the clustering procedure inevitably leads to noisy pseudo-labels. To reduce the interference of clustering noise, mutual mean-teaching (MMT) is introduced to generate reliable soft pseudo labels, however, this method is easy to fall into the local optimum. In this paper, we propose a novel Attention Random Variation (ARV) module that can be integrated into the MMT framework to develop Attention Augmented Mutual Networks (AAMN). Our ARV module generates random differences between two collaborative networks under the MMT framework to avoid the networks converging to the same kind of noise. Specifically, we propose a parameter-free Random Variation module to produce differences by randomly enhancing units of feature maps, and then combine it with an attention mechanism to enlarge networks differences and complementarity. Experimental results show that our AAMN method improves mAP of baseline method by 1.9% and 6.3% on Market-to-Duke and Duke-to-Market UDA tasks respectively.