Sta8is/Nuscenes_normals

nuScenes surface normal labels (Lotus)

Surface normal labels for the front camera (CAM_FRONT) of nuScenes, used by DINO-Foresight to evaluate the future prediction of surface normals on nuScenes. The normals are generated with Lotus (normal regression model), in the same way as the Cityscapes normal labels of DINO-Foresight. The dataset contains only the labels, not the nuScenes images.

Content

Labels for 4,014 front-camera frames (643 keyframes in samples and 3,371 intermediate frames in sweeps): 300 frames of the validation scenes and 3,714 frames of the training scenes. This includes the evaluation frame (frame 28 of the scene) of every one of the 700 training and 150 validation scenes.

Each label is a NumPy array (.npy) of shape (896, 1600, 3) and type float32, with values in [0, 1]. Its file name is the one of the corresponding nuScenes image, with the extension .npy:

nuscenes_normals
│
├───samples
│   └───CAM_FRONT          # <scene>__CAM_FRONT__<timestamp>.npy
├───sweeps
│   └───CAM_FRONT

The dataset is about 65 GB.

Download

hf download Sta8is/Nuscenes_normals --repo-type dataset --local-dir NuScenes_normals

Place the folder next to the nuScenes folder (NuScenes and NuScenes_normals), or pass its location with --normals_path in DINO-Foresight. See the DINO-Foresight repository for the nuScenes setup.

License

The labels are derived from nuScenes images. The nuScenes terms of use (CC BY-NC-SA 4.0, non-commercial use) apply.

Citation

If you use this dataset, please cite DINO-Foresight and nuScenes.

@inproceedings{karypidis2025dinoforesight,
title={{DINO}-Foresight: Looking into the Future with {DINO}},
author={Efstathios Karypidis and Ioannis Kakogeorgiou and Spyros Gidaris and Nikos Komodakis},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://arxiv.org/abs/2412.11673}
}