Sta8is/cityscapes_lotus_normals

Cityscapes Surface Normals (Lotus)

Monocular surface normal maps for the Cityscapes fine-annotated leftImg8bit frames, predicted with Lotus (a diffusion-based dense prediction model).

Each normal map corresponds 1:1 to an image in the Cityscapes gtFine/leftImg8bit fine-annotation split. This repository contains only the predicted normal maps, not the RGB images — pair them with your own copy of leftImg8bit from the official Cityscapes download page.

Note: only train and val are included. The test split is excluded (Cityscapes does not release ground truth for test, and it was not needed for this release).

Dataset structure

Normal maps are provided as a single zip archive mirroring the original Cityscapes directory layout:

leftImg8bit_normals.zip
└── leftImg8bit_normals/
    ├── train/
    │   └── <city>/
    │       └── <city>_<seq>_<frame>_leftImg8bit.npy
    └── val/
        └── <city>/
            └── <city>_<seq>_<frame>_leftImg8bit.npy

• train: 18 cities — 2,975 images
• val: 3 cities (frankfurt, lindau, munster) — 500 images
• Total: 3,475 normal maps (~82 GB uncompressed)

File names follow the Cityscapes convention, e.g.:

aachen_000000_000019_leftImg8bit.npy

which corresponds to aachen_000000_000019_leftImg8bit.png in the original Cityscapes leftImg8bit package.

Array format

• .npy files, shape (1024, 2048, 3), dtype float32
• Values are in [0, 1], the standard normal-map visualization encoding of the (x, y, z) surface normal components rescaled from [-1, 1] via (n + 1) / 2 — recover direction via n = value * 2 - 1. Note that n is the model's raw predicted direction and is not guaranteed to be unit-norm (its magnitude tends to vary with the model's per-pixel confidence, e.g. lower in flat/sky/textureless regions) — normalize (n / ||n||) if you need strict unit normals

How this was generated

Normal maps were produced by running Lotus over every fine-annotated Cityscapes leftImg8bit (train + val) RGB image.

Usage

Download and extract the archive:

huggingface-cli download Sta8is/cityscapes_lotus_normals leftImg8bit_normals.zip --repo-type dataset --local-dir .
unzip leftImg8bit_normals.zip

Used in

This dataset was used in:

• DINO-Foresight: Looking into the Future with DINO

License

This is a derivative product built on top of the Cityscapes Dataset. Use of this data is subject to the Cityscapes Dataset License Agreement — in particular it is restricted to non-commercial purposes (academic research, teaching, personal use). You must independently obtain rights to the original Cityscapes RGB imagery to make use of this data.

Citation

If you use this data, please cite Cityscapes, Lotus, and the paper that produced this dataset:

@inproceedings{Cordts2016Cityscapes,
  title={The Cityscapes Dataset for Semantic Urban Scene Understanding},
  author={Cordts, Marius and Omran, Mohamed and Ramos, Sebastian and Rehfeld, Timo and Enzweiler, Markus and Benenson, Rodrigo and Franke, Uwe and Roth, Stefan and Schiele, Bernt},
  booktitle={Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2016}
}

@article{he2024lotus,
  title={Lotus: Diffusion-based Visual Foundation Model for High-quality Dense Prediction},
  author={He, Jing and Li, Haodong and Yin, Wei and Liang, Yixun and Li, Leheng and Zhou, Kaiqiang and Zhang, Hongbo and Liu, Bingbing and Chen, Ying-Cong},
  journal={arXiv:2409.18124},
  year={2024}
}

@article{karypidis2024dinoforesight,
  title={DINO-Foresight: Looking into the Future with DINO},
  author={Karypidis, Efstathios and Kakogeorgiou, Ioannis and Gidaris, Spyros and Komodakis, Nikos},
  journal={arXiv preprint arXiv:2412.11673},
  year={2024}
}