Learning from self-discrepancy via multiple co-teaching for cross-domain person re-identification

Employing clustering strategy to assign unlabeled target images with pseudo labels has become a trend for person re-identification (re-ID) algorithms in domain adaptation. A potential limitation of these clustering-based methods is that they always tend introduce noisy labels, which will undoubtedly hamper the performance our re-ID system. To handle this limitation, an intuitive solution utilize collaborative training purify label quality. However, there exists challenge complementarity two networks, inevitably share high similarity, becomes weakened gradually as process goes on; worse still, approaches typically ignore consider self-discrepancy intra-class relations. address issue, paper, we propose multiple co-teaching framework adaptive re-ID, opening up promising direction about problem under unsupervised condition. On top that, mean-teaching mechanism leveraged enlarge difference and discover more complementary features domain. Comprehensive experiments conducted on several large-scale datasets show method achieves competitive compared state-of-the-arts.