For our dataset, we tried to collect data that includes accidents as well as diverse conditions described below. For data acquisition, we used YouTube to collect videos that have a Creative Commons License. For deduplication, first we split the videos into scenes, and then we used fingerprinting methods. Lastly, we created annotations on some of the scenes using Video Annotation Tool from Irvine, California (VATIC) software. The goal was to annotate every instance for each class: ‘cyclist’, ‘van’, ‘tram’, ‘car’, ‘misc’, ‘pedestrian’, ‘truck’, ‘person sitting’, and ‘dontcare’ as KITTI has done.