A Dataset for Object Detection in Night Rainy Scenes

Object detection is a fundamental task in computer vision. As the datasets for object detection are constructed, deep learning approaches are leveraged to show promising results. However, most of the existing datasets are aimed at normal weather scenarios, which limits the use of the trained neural networks in normal conditions. Recently, some datasets cover the night samples or rainy samples though, rare samples are utilized to focus on the night rainy scenarios. As we evaluate, the night effects and rainy effects are not decoupled. They result in a more complicated problem, thus a need for a night rainy dataset is raised to boost the deep learning tools in the data aspect. Besides, we collect the rainfall rate with a rain gauge to measure the rain intensity, splitting the dataset according to the rain intensity so that the networks can be trained according to different rainfall rates. Our whole dataset contains more than 3000 samples, with 4865 objects labeled. Specifically, the dataset contains 5 classes labeled as pedestrian, car, bus, truck, and motor cyclist. To show the effectiveness of the dataset, we use the latest detection algorithm trained on the open dataset and our dataset for comparison. The significant improvement reveals that our dataset are unique and strongly pushes the networks to the practical useful end.