IDM: An Intermediate Domain Module for Domain Adaptive Person Re-ID

Illustration of IDM model

Abstract

This paper proposes a plug-and-play Intermediate Domain Module (IDM) to tackle the problem of UDA re-ID. From a novel perspective that intermediate domains can bridge the source and target domains, our purposed IDM module can generate appropriate intermediate domain representations to better transfer the source knowledge to improve the model’s discriminability on the target domain. The intermediate domains’ distribution is controlled by the two domain factors generated by the IDM module. Specifically, we propose the bridge losses to enforce the intermediate domains to be located onto the appropriate path between the source and target domains in a manifold. Besides, we also propose a diversity loss to constrain the domain factors to prevent the intermediate domains from being over-fitting to either of the source and target domains. Extensive experiments have shown the effectiveness of our method.

Publication
International Conference on Computer Vision

Related