NACrash: A Densely Annotated North American Crash Video Dataset and Benchmark for Multimodal Large Language Models

Kaidi Liang, Ke Li, Debojyoti Biswas, Yuxin Ding, Xianbiao Hu, Ruwen Qin

WACV 2027 (Evaluations and Datasets Track) | GitHub

NACrash is a richly annotated dashcam crash video dataset collected from North American traffic environments. It contains 4,024 video clips (about 2.55 million frames and 23.6 hours of footage) with dense natural-language descriptions, multi-attribute annotations based on standards-aligned crash taxonomies, and 33,920 question-answering (QA) pairs over nine crash video understanding tasks. NACrash-Bench is a balanced benchmark of 1,000 videos and 8,700 QA pairs.

Contents

nexar_videos/          # 1,458 Nexar-sourced video clips (nexar_XXXX.mp4)
LICENSE_Nexar          # Nexar Open Data License (applies to the videos in nexar_videos/)

The annotations (per-video attributes and QA pairs for NACrash and NACrash-Bench) are available in the GitHub repository. The video_path field in the annotation files refers to the file names in nexar_videos/.

The 2,566 YouTube-sourced clips are not redistributed here. The annotation spreadsheet lists the source channel of each YouTube clip.

License

• Videos in nexar_videos/ are derived from the Nexar Collision Prediction Dataset and are redistributed under the Nexar Open Data License, included in this repository as LICENSE_Nexar. Copyright (c) 2025 Nexar Inc. The license prohibits resale, reidentification of people or vehicles, and other uses listed in the license. Please read it before using the videos.
• NACrash-created artifacts (annotations, QA pairs, and benchmark metadata) are released under the CC BY-NC 4.0 license.

Citation

If you use NACrash, please cite our paper:

@inproceedings{liang2027nacrash,
  title     = {NACrash: A Densely Annotated North American Crash Video Dataset and Benchmark for Multimodal Large Language Models},
  author    = {Liang, Kaidi and Li, Ke and Biswas, Debojyoti and Ding, Yuxin and Hu, Xianbiao and Qin, Ruwen},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}

As required by the Nexar Open Data License, please also cite the source of the Nexar-sourced videos:

Moura, Daniel C., and Zvitia, Orly. "Nexar Collison Dataset." Hugging Face, 2025, https://huggingface.co/datasets/nexar-ai/nexar_collision_prediction.