Multi-intermediate Feature with Multi-stage Fusion for Domain Adaptive Person Re-ID

Abstract

Unsupervised domain adaptive person re-identification (UDA Re-ID) aims at obtaining more robust and discriminative feature on unlabeled target domain by transferring knowledge from labeled source domain. Especially, approaches based on intermediate domains mixing source and target domain have achieved the state-of-the-art results for UDA Re-ID task. Specifically, we design a novel method which define multiple intermediate domain features which is a fusion of source domain and target domain and play critical bridging role with a better path from source to target domain. Furthermore, the multi-intermediate domain’s representation consists of transfer representation from the previous fusion stage and the multi-stage hidden representations captured from source domain and target domain. In order to obtain better appropriate multi-intermediate domains, we define a multi-node bridge losses on the multi-intermediate domains, source domain and target domain and it can guide to obtain a more appropriate path with multiple node defined as multi-intermediate feature from source domain to target domain. Our proposed method achieves better performance comparing with the state-of-the-art approaches in common UDA Re-ID tasks, and the mAP gain is up to on the challenging DukeMTMC → market1501 benchmark.