Cluster-Match Collaborative Learning for Unsupervised Heterogeneous Person Re-Identification

Heterogeneous person re-identification (ReID) aims to match different captured images of the same individual in challenging situations where the images exhibit strong differences, for instance, visible-infrared ReID or cross-resolution ReID. Existing unsupervised approaches for heterogeneous ReID, typically employ clustering or matching algorithms to associate heterogeneous images of the same person (aka heterogeneous positive pairs), thereby forming mixed clusters that facilitate model optimization. However, clustering and matching algorithms struggle to simultaneously guarantee both the quantity and the reliability of mixed clusters, which handicaps learning. To address this issue, we propose a cluster-match collaborative learning method (CMCL). First, we introduce a covariance-level heterogeneous alignment module, which generates whitening and coloring matrices from covariance statistics and aligns heterogeneous image features by whitening and re-coloring each feature. This process encourages heterogeneous positive pairs to be grouped into the same cluster. Subsequently, we develop a cluster-match positive mining module that first performs clustering on heterogeneous images to produce reliable mixed clusters, and then applies a matching algorithm to ensure a high number of these clusters. In the optimization stage, in addition to the conventional centroid-level homogeneous and heterogeneous contrastive losses, we introduce an instance-level mixed contrastive loss, which leverages the mixed clusters generated by clustering to promote identity relevance and heterogeneous invariance at the instance level. We further design a heterogeneous association sampling strategy to ensure that images within each minibatch form online contrastive pairs, thereby providing a more accurate optimization direction. Experimental results on multiple visible-infrared and cross-resolution datasets demonstrate the superiority and generalization ability of the proposed CMCL methodology. It not only surpasses existing unsupervised approaches but also achieves performance competitive with some supervised methods.