Unsupervised Person Re-Identification via Multi-Order Cross-View Graph Adversarial Network

Abstract

Unsupervised person re-identification (re-id) is an effective analysis for video surveillance in practice, which can train a pedestrian matching model without any annotations, and it is easy to deploy in unseen camera scenarios. The most challenging problem in unsupervised re-id task is the huge distribution-gap among different camera views, and the intrinsic correlations in unlabeled identities are also complicated to sufficiently explored. This paper proposes a Multi-order Cross-view Graph adversarial Network (MCGN) to bridge the cross-view distribution-gap, and mine the inherent discriminative information by multi-order triplet correlations. Specifically, MCGN firstly exploits graph representations by a cross-view graph convolutional network according to intra-view and inter-view graph structure, and then encodes each pedestrian image into a view-shared feature space, which is iteratively trained by a graph generative adversarial learning strategy to deeply bridge the distribution-gap. Finally, this paper proposes a multi-order discriminative learning module for composing reasonable triplet samples according to multi-order similarity correlations among unlabeled pedestrian images. Furthermore, sufficient experiments are conducted in two large scale person re-id datasets (Market-1501 and DukeMTMC-reID). The comparison to state-of-the-art methods and ablation study demonstrate the superiority of MCGN and the contribution of each module proposed in this paper.