We have constructed a novel benchmark, Challenge Of Out-Of-Label (COOOL) in Autonomous Driving. This benchmark features high-resolution videos from real-world driving scenarios. COOOL is an evaluation benchmark focused exclusively on diverse roadway hazards, including exotic animals like kangaroos and wild boars, inanimate and hard-to-predict hazards like plastic bags and smoke, and standard hazards like cars and pedestrians. Using vision information to detect out-of-label hazards on roadways is an often overlooked problem in autonomous driving research; our benchmark highlights this issue exclusively. It includes detailed labels for each object in every frame, enhancing research capabilities. A unique aspect of our benchmark is the “Tag” information for each frame, which provides insights into the car's movements and the driver's decisions. Undergraduate students at the University of Colorado Colorado Springs (UCCS) and local high school students have annotated the benchmark using a professional computer vision annotation platform under the guidance of expert graduate students and professors.

[WACV 2025]  [COOOL]