Divide-and-Regroup Clustering for Domain Adaptive Person Re-identification

Clustering is important for domain adaptive person re-identification(re-ID). A majority of unsupervised adaptation (UDA) methods conduct clustering on the target and then use generated pseudo labels training. Albeit important, pipeline adopted by current literature quite standard lacks consideration two characteristics re-ID, i.e., 1) a single has various feature distribution in multiple cameras. 2) person’s occurrence same camera are usually temporally continuous. We argue that multi-camera hinders because it enlarges intra-class distances. In contrast, temporal continuity prior beneficial, offers clue distinguishing some look-alike (who far away from each other). These insight motivate us to propose novel Divide-And-Regroup (DARC) re-ID UDA. Specifically, DARC divides unlabeled data into camera-specific groups conducts local within camera. Afterwards, regroups those clusters potentially belonging unity. Through this divide-and-regroup pipeline, avoids directly across cameras focuses individual Moreover, during clustering, uses distinguish thus reduces false positive labels. Consequentially, effectively errors improves Importantly, we show compatible many label-based UDA brings general improvement. Based recent method, advances state art (e.g, 85.1% mAP MSMT-to-Market 83.1% PersonX-to-Market).