Augmentation-Based Object Detection for Winter Time Applications

Despite the significant advancements in object detection, there is still a number of detection tasks that are challenging and difficult. To eliminate them and to push forward the research in this area a technological "IceVision" Challenge was organized in Russia in 2019. The main objective of this Challenge is to develop new approaches for object detection in severe weather conditions focusing on winter time. The obtained results are supposed to be implemented for autonomous driving. This paper reports the approach that was used by our team in this competition. The approach is based on a Cascade R-CNN model in combination with data augmentation. It allowed us to take fourth place among 187 participants on the preliminary stage of the completion.