Pseudo-label Based Unsupervised Momentum Representation Learning for Multi-domain Image Retrieval

Although many current cross-domain image retrieval researches have made good progress, most of the works is targeted at specific domains. At the same time, we also noticed that many works are based on manually annotated images. In this paper, in order to solve the above problems, we propose a new paradigm applied to multi-domain image retrieval. The specific solution is as follows: 1) The momentum contrastive mechanism is used to deeply mine the semantic features shared by images of the same category in the same domain. 2) Pseudo-labels are learned through clustering learning to provide supervised information for contrastive learning. 3) By aligning images of the same category in different domains, the aggregation of common features across multiple domains is achieved. Experiments on multi-domain datasets demonstrate the feasibility of our idea. Code will be made public in https://github.com/DannielGe/PUMR .