In spite of the rapid expansion and investment into autonomous driving systems, dealing with rare or novel roadway hazards remains fundamental obstacle to the safe deployment of these systems. Foundation models and LLMs have made significant progress in recognizing a wide variety of semantic relations from images. However, utilizing these systems with unseen or rare data distributions (ie images taken from a moving vehicle), remains challenging. Additionally, novel, damaged, and distant/low-resolution hazards present their own challenge, as misclassifying these hazards as a nuisance could prove fatal in deployment.

We've collected ~200 high-quality dashcam videos with a variety of hazards and accompanying roadway objects.
Many videos contain more than one hazard, with some changing over time..
This challenge requires participants to address three tasks:
- When has the driver reacted to a hazard.
- Which bounding box(es) contain hazard(s).
- Describe what the hazard(s) are.