Decouple and Reconstruct: Mining Discriminative Features for Cross-domain Object Detection

In recent years, a great progress has been witnessed for cross-domain object detection. Most state-of-the-art methods strive to handle the relation between local regions by calibrating cross-channel and spatial information to enable better alignment. They succeed in improving the generalization of the model, but implicitly drive networks to pay more attention on the shared attributes and ignore the domain-specific feature, which limits the performance of the algorithm. In order to search for the equilibrium between transferability and discriminability, we propose a novel adaptation framework for cross-domain object detection. Specifically, we adopt a style-aware feature fusion method and design two plug-and-play feature component regularization modules, which repositions the focus of the model on domain-specific features by restructuring the style and content of features. Our key insight is that while it is difficult to extract discriminative features in target domain, it is feasible to assign the underlying details to the model via feature style transfer. Without bells and whistles, our method significantly boosts the performance of existing Domain Adaptive Faster R-CNN detectors, and achieves state-of-the-art results on several benchmark datasets for cross-domain object detection.