Event
Proposes end-to-end driving image degradation benchmark DriveDegrade
1 reports · 1 sources
Haohua Que et al. propose DriveDegrade, an image degradation tolerance benchmark for pure-camera end-to-end (E2E) driving: it injects 16 types of corruption at 5 severity levels online in the image loader, covers 15 strategies with a single operator, evaluates open-loop planning on nuScenes and NAVSIM, and includes a CARLA closed-loop anchor. The paper has been accepted to WACV 2027. The paper reports that a released vision-language-action (VLA) planner performs flatly along these two axes, and masking all six of its cameras causes it to lose only 11.5%.
Report timeline
- DriveDegrade: An Image-Degradation Tolerance Benchmark for Camera-Only End-to-End Driving Planning
The paper proposes DriveDegrade, an image-degradation tolerance benchmark for camera-only end-to-end (E2E) driving. It injects 16 corruption types at 5 severity levels online in the image loader, uses a single operator to cover 15 policies, and evaluates open-loop planning on nuScenes and NAVSIM, with a CARLA closed-loop anchor.
arXiv cs.RO ·