KNOWLEDGE DISTILLATION FOR SEMANTIC SEGMENTATION APPLIED TO AUTONOMOUS DRIVING IN ADVERSE WEATHER CONDITION

Semantic segmentation is a fundamental perception task for autonomous driving system, lot of work compare their result in a clear weather in daylight scenario using Cityscapes dataset, however one of the main challenge of autonomous driving nowadays is to build systems that are robust to adverse weather. Another important factor to take into ac- count when it comes to semantic segmentation is that most state-of-the-art architectures are not designed to run in real-time on an embedded system that can also have limited memory. For this reason, this work first evaluates real-time architectures on ACDC, a popular adverse weather dataset, then after selecting one of the best performing model on this dataset, proposes to improve the real-time network using different teacher-student knowledge distillation techniques. The experiments first highlight several improvements in term of model compression using a network with reduced number of parameters without significant performance loss. We then conduct other experiments using a network capable of low-latency inference obtaining a distilled model that perform well on ACDC.