Domain Adaption Object Detection with Global - Local Contrastive Learning and Co-training Network
Ming Zhao, Xing Wei, Yang Lu, Ting Bai, Chong Zhao · 2022
With the continuous enrichment and development of technical means, object detection has developed rapidly. However, object detection relies on a plenty of number of labeled datasets, and when applied to unlabeled or less-labeled datasets, the performance suffers greatly due to the effect of domain transfer. To solve this problem, we propose a contrastive learning domain adaptation method, which transfers from the labeled source domain to the unlabeled target domain. In our paper, we first propose a contrastive learning network for global and local matching, where the global contrastive module is used to learn image-level representations globally, while the contrastive learning module for local matching is beneficial for semantic segmentation. Local region representation can be better learned. Then, both domain samples are fed into the co-training network at the same time, so that they can share weights with each other. The design of the two modules promotes global and local consistent representation while improving detection accuracy. Experimental results show that our proposed model framework improves on datasets, and existing methods are more efficient and competitive.