Clouded Judgments: An Empirical Analysis of Vision-Based Object Detection Systems in Harsh Weather Conditions

At the core of environment understanding and interpretation, there is the capability of detecting and recognising the agents that live and interact in that environment, at all times. Such capability is crucial for road monitoring and traffic analysis, as well as to support autonomous mobility. Indeed, to be safe, driverless vehicles need to rely on accurate and robust perception models. While, in most cases, current systems showcase a reliable and accurate behaviour, adverse weather conditions, such as rain, fog, and snow, still pose undeniable challenges to object detection. In this paper, we conduct a comprehensive analysis of how these environmental factors degrade visual perception performance across different detection models. By dissecting detection failure modes under various weather scenarios, we aim to expose critical limitations and performance inconsistencies. In particular, we investigate the impact of weather-tuned training strategies on detection performance; differently from similar works in the area, we focus on real and not synthetic images. Our analysis identifies key performance gaps and recurring limitations that hinder reliable detection and provides valuable insights towards the development of robust visual perception systems.