Learning to Remove Bad Weather: Towards Robust Visual Perception for Self-Driving

Abstract

Visual perception plays a vital role in generating the intelligent actions of autonomous vehicles. However, various bad weather degradations can create visibility problems that impair the performance of high-level tasks. Though image enhancement techniques have been extensively studied for safe self-driving in bad weather, few studies have dealt with both enhancement and perception simultaneously. To cope with bad weather conditions for high-level perception, we propose an end-to-end deep learning-based framework, which connects the enhancement network with the perception network. To this end, we design a universal enhancement network that can address multiple bad weather conditions and assist in producing promising perception results. To this end, we design a universal enhancement network that can address multiple bad weather conditions and assist in producing promising perception results. To this end, we design a universal enhancement network that can address multiple bad weather conditions and assist in producing promising perception results. IEEE