For the convenience of participants, we provide a working baseline using simple methods to address each task. For driving state, we estimate the distance between bounding boxes over frames, and use Linear Regression to detect slow-downs. For Hazard detection, we simply assume the bounding box closest to the center of the image is the hazard and caption them using clip interrogator.

This is a very simple baseline, there is significant room for improvement. For example, computing hazard trajectories, using better captioning models, car speed and orientation detection could all improve results.

We provide the baseline on our GitHub: https://github.com/alshami52/COOOL_benchmark