Road Anomaly

This dataset contains images of unusual dangers which can be encountered by a vehicle on the road – animals, rocks, traffic cones and other obstacles. Its purpose is testing autonomous driving perception algorithms in rare but safety-critical circumstances.

The dataset contains images with associated per-pixel labels. The labeling has been performed with our LabelGrab tool.

Most frames retain the editor’s files and can be further edited; some are missing because the labeling was done at a different resolution and rescaled. If you need the editor files and instance labels for all frames, please contact us.

The images are provided for research purposes. We do not own the images; the authors and sources are listed in credits.txt.

For any additional information, please contact Krzysztof Lis.

This dataset is now part of the expanded Segment Me If You Can benchmark.


Directory structure


  • frame_list.json – list of all image file names
  • credits.txt – sources of the images
  • frames/
    • frame_name.webp – image
    • frame_name.labels/labels_semantic.png – per pixel label, background = 0 and anomaly = 2
    • frame_name.labels/labels_instance.png – instance id of objects, some are missing at this time.


Detecting the Unexpected via Image Resynthesis

K. M. Lis; K. K. Nakka; P. Fua; M. Salzmann 

2019-10-27. IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, SOUTH KOREA, Oct 27-Nov 02, 2019. p. 2152-2161. DOI : 10.1109/ICCV.2019.00224.