Region Feature Disentanglement for Domain Adaptive Object Detection

Abstract

In recent years, deep learning based object detection has shown impressive results. However, applying an object detector learned from one data domain to another one often faces performance degradation due to the domain shift problem. To improve the generalization ability of object detectors, the majority of existing domain adaptation methods alleviate the domain bias either on the feature encoder or instance classifier by adversarial learning. Differently, we try to alleviate domain discrepancy in the region proposal network (RPN) by performing feature disentanglement. To this end, an extractor is devised to extract domain-specific foreground representations from both the source and target features, respectively. Then, domain-invariant representations are decomposed from the domain-specific features by the disentanglement module. Through the decoupling operation, the gap between the domain-specific and domain-invariant features is enlarged, which promotes RPN feature to contain more domain-invariant information. Furthermore, we propose dynamic weighted adversarial training to alleviate the unstable training caused by adversarial learning. We conduct extensive experiments on multiple domain adaptation scenarios, and our experiment results demonstrate the effectiveness of our proposed approach.