We introduce a comprehensive public dataset, NightOwls, for pedestrian detection at night. In comparison to daytime conditions, pedestrian detection at night is more challenging due to variable and low illumination, reflections, blur, and changing contrast.
NightOwls consists of 279k frames in 40 sequences recorded at night across 3 countries by an industry-standard camera, including different seasons and weather conditions. All the frames are fully annotated and contain additional object attributes such as occlusion, pose and difficulty, as well as tracking information to identify the same object across multiple frames. A large number of background frames for evaluating the robustness of detectors is included, a validation set for local hyper-parameter tuning, as well as a testing set for central evaluation on a submission server is provided.
As a baseline for pedestrian detection at night time, we compare the performance of ACF, Checkerboards, Faster R-CNN, RPN+BF, and SDS-RCNN. In particular, we demonstrate that state-of-the-art pedestrian detectors do not perform well at night, even when specifically trained on night data, and we show there is a clear gap in accuracy between day and night detections. We believe that the availability of a comprehensive night dataset may further advance the research of pedestrian detection, as well as object detection and tracking at night in general.
