I Had a Bad Day: Challenges of Object Detection in Bad Visibility Conditions

Automated vehicles must be able to correctly perceive the environment in every conceivable situation with the help of their sensor stack. The resulting data streams are typically processed by algorithms that detect and classify objects in the scene. In order to ensure safe driving, these algorithms are expected to perform with sufficient accuracy even in the harshest weather conditions, such as rain, fog and snow. However, this is not the case for the current generation of object detection models. A sharp drop in performance can be observed as soon as these are exposed to adverse weather situations. This can be attributed to the weak representation of training data in bad visibility conditions and detection architectures that are not designed to handle them properly. To address this problem, we propose three small scale annotated datasets that include challenging adverse weather conditions. We evaluate state-of-the art object detection models on a variety of datasets and quantify the performance drop for each algorithm in adverse weather conditions. To this end, we show that current object detection models suffer from severe performance loss due to adverse weather effects and identify common challenges of object detection in bad visibility conditions.