Let it Snow: On the Synthesis of Adverse Weather Image Data

Camera systems of automated vehicles capture images from the surrounding environment and process these datastreams with algorithms to detect and classify objects. A lot of research has been devoted to improve object detection algorithms in order to provide highly accurate detection results in real time. However, these algorithms show a strong drop in performance as soon as they are exposed to adverse weather. Poor weather conditions such as rain, fog or snow lead to a reduction in visibility and thus objects are more difficult to recognize or not visible at all. This leads to a high degree of uncertainty for an automotive camera system. To enable automated driving, camera systems must be able to cope with adverse weather and the associated high uncertainty. Including more weather image data when training the algorithms can improve object detection in bad visibility conditions. However, weather image data is difficult to collect in reality and thus only available to a limited extent. In this work, we evaluate the possibility of using Generative Adversarial Networks to create synthetic weather image data. For this purpose, we compare the generated images of different network architectures trained on a diverse weather dataset collected from Flickr. The resulting data is evaluated qualitatively and quantitatively with respect to its realism and suggests that our approach is capable of generating realistic weather images.