A Clearer Image: Improving Object Detection in Real Rainy Conditions with Two-Stage Processing
Mingxuan Yang, Xiyu Han, Xianyao Ping, Zipeng Li, Jing Xiao · 2023
Object detection is one of the most important and challenging branches of computer vision, which has been widely applied in social life, such as intelligent security, autonomous driving, and so on. However, the performance of object detection could degrade rapidly under challenging weather scenarios including rainy conditions. Despite existing technologies having made impressive advancements, they are almost under synthetic rainy-clean pairs, which have a huge domain gap between synthetic and real rain. In this work, we propose an unsupervised deraining method under real rain and utilize a non-local contrastive learning for better decoupling the rain layer from the clean image. Furthermore, we introduce a semi-supervised detector during detection module to reduce the inconsistency of pseudo labels for better facilitating contrastive learning. Extensive experiments with different deraining and object detection methods demonstrate that our two-stage method obtains competitive performance in complex weather scenarios.