We selected challenging scenes from the public NuScenes, Waymo, and KITTI dataset for scenario editing. These included scenes with over 50 vehicles, data from different countries, and varied weather (sunny, rainy) and lighting conditions (daytime, nighttime). Every tenth frame of the driving videos was used as a test frame.
| Dataset | Scene IDs |
|---|---|
| NuScenes |
scene-0003, scene-0105, scene-0127, scene-0399, scene-0405, scene-0443, scene-0514, scene-1098 |
| Waymo |
seg109239..., seg125050..., seg113555..., seg128560... |
| KITTI |
2011_09_26_drive_0014.sync, 2011_09_26_drive_0017.sync, 2011_09_26_drive_0018.sync |
Beyond photorealism, synthetic data for autonomous driving systems must also exhibit algorithmic realism. Following the real-world corner case CODA dataset setup, we selected 8 safety-critical scenarios from the public NuScenes for evaluation.