Rethinking all-in-one adverse weather removal for object detection

Despite significant progress has been made in image restoration under adverse weather conditions, these methods primarily focus on the quality of image reconstruction, leaving their impact on downstream object detection unknown. In this paper, we rethink all-in-one adverse weather removal for object detection. Specifically, we contribute the first multi-weather image restoration dataset tailored for autonomous driving scenarios, comprising 6000 image pairs along with object detection labels, named Multi-Weather6k. Based on this dataset, we conduct a benchmark study on existing methods for joint image restoration and object detection. Furthermore, we develop an effective MLP-based all-in-one image de-weathering framework to better solve this task. The proposed architecture consists of the feature mixing block and the feature prompt block. The former enhances feature modeling by exploiting global and local correlations, while the latter guides image restoration by modulating multi-degraded features using prompt learning. Experimental results show that our proposed method not only achieves consistently superior restoration performance across various weather degradation scenarios but also yields improved object detection results, outperforming the state-of-the-art approaches. The dataset will be available to the public.