Reproducing `scenario_labels` / `camera_labels`

[ayane11]

Hi, thanks for the release.

The training pipeline reads labels/scenario_labels and labels/camera_labels from the DriveMoE Hugging Face repository. However, I could not find the code that generates these labels from the original Bench2Drive annotations.

Could you share how these labels were generated? Specifically:

• Are they automatically derived from Bench2Drive annotations?
• Is the label-generation script available?
• How are Bench2Drive scenarios mapped to the 7 DriveMoE scenario categories?
• How is the per-frame camera label selected?

Thanks!

[jayyoung0802]

Hi @ayane11,
Regarding [Question 1-2], we automatically derived from Bench2Drive annotations—the annotations generated by Bench2Drive-VL in the offline dataset.

Regarding [Question 3], the 7 scenario categories are defined by us, and there's no special design, just the most intuitive driving understanding.

Regarding [Question 4], the per-frame camera label is selected by scenarios. We use a set of rules to identify special cases using the annotation information in the Bench2Drive dataset:

• When the ego vehicle needs to turn at an intersection. In this case, the ego vehicle’s is_in_junction flag is true, and the current command is left turn or right turn. We label the front-side camera that is close to the target exit.
• When the ego vehicle needs to change lanes. This happens when the current command is left lane change or right lane change, when there is an obstacle within a certain distance in the ego lane, or when the ego vehicle is not in the target lane. If the target lane has the same driving direction as the ego vehicle, we label the corresponding rear-side camera. Otherwise, the ego vehicle needs to use the opposite lane, so we label the front-side camera. In particular, after the ego vehicle passes an obstacle and needs to change back to the original lane, we label the front-side camera facing the original lane.
• In highway merging scenes. When the current scenario is highway merging or cut-in, we use the ego vehicle’s lane position and distance to the junction to decide where the merging point is. Then we label the side camera facing the merging point.
• When avoiding an emergency vehicle. If there is a fast-moving emergency vehicle in the scene, the ego vehicle needs to avoid it. Therefore, we label the camera that sees the emergency vehicle.