Unsupervised Domain Adaptation for Object Detection using Distribution Matching in Various Feature Level

Abstract

As the research on deep learning has become more active, the need for a lot of data has emerged. However, there are limitations in acquiring real data such as digital forensics, so domain adaptation technology is required to overcome this problem. This paper considers distribution matching in various feature level for unsupervised domain adaptation for object detection with a single stage detector. The object detection task assumes that training and test data are drawn from the same distribution; however, in a real environment, there is a domain gap between training and test data which leads to degrading performance significantly. Therefore, we aim to learn a model to generalize well in target domain of object detection by using maximum mean discrepancy (MMD) in various feature levels. We adjust MMD based on single shot multibox detector (SSD) model which is a single stage detector that learns to localize objects with various size using a multi-layer design of bounding box regression and infers object class simultaneously. The MMD loss on high-level features between source and target domain effectively reduces the domain discrepancy to learn a domain-invariant feature in SSD model. We evaluate the approaches using Syn2real object detection dataset. Experimental results show that reducing the domain shift in high-level features improves the cross-domain robustness of object detection, and domain adaptation works better with simple MMD method than complex method as GAN.