Joint Feature-level and Pixel-level Domain Adaption for Object Detection in the Wild

The obeject detector trained on a source dataset probablely fails to be generalized well to a target dataset due to the domian shift between different distributions. This paper aims at reducing such domain gap without any available annotation of the target domain. It is achieved through a novel unsupervised domain adaption algorithm which improves the robustness of Faster R-CNN on two levels: 1) on the feature level, a domain discriminator is appended to the feature extractor of Faster R-CNN; 2) on the pixel level, multiple feature maps from the feature extractor are connected to a generator which feeds the fused features in a form of pictures to a later domain discriminator. When used alone, either one can improve the performance in a adversarial training manner, the former acheives a high precison-rate while the latter gains a high recall-rate. While applying both components can play the role of complementary advantages. After domain adaption, the accuracy of Faster R-CNN on the publicly used Virtual KITTI dataset increases from 51.2% to 69.5%, the accuracy on the publicly used Foggy Cityscapes dataset increases from 20.2 % to 30.3 % which outperforms the state-of-the-art method by 1.3 % in the accuracy gain.