Multiformer-based hybrid learning with outlier re-assignment for unsupervised person re-identification

Min Jiang, Qian Zhang, Jun Kong

Abstract: Recent studies have shown the effectiveness of generating pseudo-labels by measuring feature similarity in unsupervised person re-identification (ReID). However, most of these methods ignore the distribution discrepancy among cameras, which leads to degraded accuracy in generated pseudo-labels. Besides, limited by the small receptive field and downsampling operations, the convolution-based schemes cannot effectively capture fine-grained information and global dependencies. In this paper, we propose multiformer-based hybrid learning with outlier re-assignment (MHOR) for unsupervised person ReID. Firstly, to mitigate the distribution discrepancy among cameras, we design a multi-branch transformer (Multiformer) network, including inter-camera transformer (Interformer) and intra-camera transformer (Intraformer). The Multiformer network allows the computation of feature similarity to be performed separately for inter-camera and intra-camera scenarios. Secondly, to further enhance the fine-grained information and global dependencies feature representation in the context of Multiformer, a dynamic outlier re-assignment (DORA) strategy is proposed to reassign pseudo-labels by computing the affinity matrix of outlier samples with cluster centers. Thirdly, to improve the quality of pseudo-labels, we propose hybrid contrastive learning (HCL), which uses instance-level contrastive learning to distinguish different pedestrian features and introduces cluster-level contrastive learning to alleviate the unreliability problem caused by noisy labels. Finally, extensive experiments show that MHOR can significantly surpass the performance of previous works on unsupervised tasks for person ReID.