DarkDriving: A Real-World Day and Night Aligned Dataset for Autonomous Driving

Project page for DarkDriving, accepted at ICRA 2026.

Wuqi Wang1*, Haochen Yang2*, Baolu Li2, Jiaqi Sun1, Xiangmo Zhao1, Zhigang Xu1, Qing Guo3, Haigen Min1, Tianyun Zhang2, Hongkai Yu2
1Chang'an University 2Cleveland State University 3A*STAR

*Equal contribution

Paper arXiv Code & Data
DarkDriving trajectory tracking and pose matching pipeline

Abstract

DarkDriving is a real-world benchmark for low-light autonomous driving. It contains 9,538 day-night image pairs with centimeter-level alignment in both location and scene content, captured on a 69-acre closed driving test field and annotated with 2D object boxes.

Examples from the DarkDriving paired day and night dataset

Dataset

Robust LocalizationA high-precision point-cloud map and NDT localization provide stable centimeter-level vehicle poses.
Trajectory TrackingThe automated vehicle repeats consistent daytime and nighttime trajectories using closed-loop control.
Pose MatchingDay-night frames are matched by pose and refined to preserve spatial correspondence in dynamic scenes.
Road layouts and illumination configurations in DarkDriving

Results

DarkDriving enables supervised enhancement with real paired daytime targets. The main comparison summarizes both full-reference and no-reference image-quality metrics across representative low-light enhancement methods.

Quantitative low-light enhancement results on DarkDriving

BibTeX

@inproceedings{wang2026darkdriving,
  title     = {DarkDriving: A Real-World Day and Night Aligned Dataset for Autonomous Driving in the Dark Environment},
  author    = {Wang, Wuqi and Yang, Haochen and Li, Baolu and Sun, Jiaqi and Zhao, Xiangmo and Xu, Zhigang and Guo, Qing and Min, Haigen and Zhang, Tianyun and Yu, Hongkai},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year      = {2026}
}