A Joint Detection Network for Snowy Weather with Image Recovery and Domain Adaptation

With the development of domain adaptation method, object detection models obtain progress performance in inclement weather conditions. However, a large gap between systhetic datasets and real datasets could make adaptation a challenging task, which hinders the network's learning of domain-invariant features. To address such issue, we propose a novel joint framework for object detection under snowy weather. We design the Progressive Recovery Module which averages the restoration image with original image, in order to alleviate the domain shift. Further, we design a novel hierarchical domain adaptation structure that skillfully enhances the learning of domain-invariant features between two domains by fusing three feature layers. Experimental results show that our method achieves higher accuracy compared to the recent state-of-the-art methods in real-to-world snowy images.