Overview

The MM-AU dataset for ego-view accident video understanding is collected from the publicly available ego-view accident datasets, such as CCD, A3D, DoTA, and DADA-2000, and various video stream sites, such as Youtube, Bilibili, and Tencent etc. To the best of our knowledge, MM-AU is the largest and most fine-grained ego-view multi-modal accident dataset.

MM-AU is ONLY free for non-commercial use. Examples of non-commercial use include but are not limited to educational use, such as in schools, academies, universities, etc., and some research use. If you intend to use MM-AU for commercial purposes, we encourage you to contact us for obtaining commercial license by contacting us with lotvsmmau@gmail.com.

Downloads

Here, we list the different parts of MM-AU dataset. 

• Raw Videos: We provide 11,727 videos with 2,195,613 frames.

• Raw Detection: To facilitate the object-centric accident video understanding, we annotate 7 classes of road participants (i.e., cars, traffic lights, pedestrians, trucks, buses, cyclists, and motorbikes) in MM-AU. To fulfill an efficient annotation, we first employ the YOLOX detector (pre-trained on the COCO dataset), to initially detect the objects in the raw MM-AU videos to make a coarse annotation set.

• Refinement Detection: We took three months to manually correct the wrong detections using LabelImg every five frames by ten volunteers, and 2,233,683 bounding boxes within 463,207 frames are obtained. Each bounding box is double-checked for the final confirmation.

• Accident Window: The video attributes and annotated information for the accident windows have been integrated into an Excel file for the convenience of users.

• Text Description Annotation: MM-AU annotates three kinds of text descriptions: accident reason, prevention advice, and accident category descriptions, and we have integrated these text annotations into an Excel file for the convenience of users.

The download link can be obtained by registering and sending email to us lotvsmmau@gmail.com.

For streamlined accessibility, our GitHub hosts comprehensive usage instructions for each task with the MM-AU dataset.

Sign Up Login