Dark Zurich is an image dataset containing 8,779 images taken at night, dusk, and daytime, along with the GPS coordinates of the camera corresponding to each image. These GPS annotations can be used to construct correspondences with time spans, such as matching images taken at night or dusk with corresponding images taken during the day.

These properties make the Dark Zurich dataset useful for creating models and systems that perform:

• Domain adaptation (unsupervised, weakly supervised, or semi-supervised), such as semantic segmentation or object detection;
• Image translation/style transfer at different times of the day;
• Robust image matching/visual localization across multiple domains;
• Other autonomous driving or robotics related visual perception tasks.