Resources
Code for loading the data, evaluating detectors and reproducing the baselines, plus the tools used to generate and verify the dataset.
Software
- Dev kit
- The
carlanomalyPython package: dataset loaders for every modality and streaming evaluators for all four tiers. - Baselines
- Training and evaluation scripts for the baseline detectors reported on the tasks page.
- CarlaGen
- Combines Hydra configs with the CARLA Python API. Use it to generate your own dataset versions with different sensors and resolutions. Release pending.
- Labeling interface
- The web interface used to manually verify and label every scenario. Release pending.
Pretrained checkpoints
Checkpoints for the baseline models, ready to load with the baselines code.
| Baseline | Model | Checkpoint | Size |
|---|---|---|---|
| DeepLabV3-ResNet50 | Image segmentation | deeplabv3_r50_msp.pt | 168 MB |
| Mask2Former (Swin-L) | Image segmentation, scored with RbA | mask2former_swin_l_full.pt | 862 MB |
| Mask4Former | Panoptic point-cloud segmentation | mask4former.pt | 454 MB |
| FFP (U-Net) | Future-frame prediction | unet_ffp.pt | 31 MB |
| V-JEPA 2 (ViT-L) | Future-frame prediction, latent-space error | vjepa2_predictor.pt | 88 MB |