naufalso/cityscape-adverse

Cityscape‑Adverse

A benchmark for evaluating semantic segmentation robustness under realistic adverse conditions.

Overview

Cityscape‑Adverse extends the original Cityscapes dataset by applying eight realistic environmental modifications—rainy, foggy, spring, autumn, snowy, sunny, night, and dawn—using diffusion‑based image editing. All transformations preserve the original 2048×1024 semantic labels, enabling direct evaluation of model robustness in out‑of‑distribution scenarios.

Paper & Citation

Cityscape‑Adverse: Benchmarking Robustness of Semantic Segmentation with Realistic Scene Modifications via Diffusion‑Based Image Editing
N. Suryanto, A. A. Adiputra, A. Y. Kadiptya, T. T.-H. Le, D. Pratama, Y. Kim, and H. Kim, IEEE Access, 2025.
DOI: 10.1109/ACCESS.2025.3537981

@ARTICLE{10870179,
  author={Suryanto, Naufal and Adiputra, Andro Aprila and Kadiptya, Ahmada Yusril and Le, Thi-Thu-Huong and Pratama, Derry and Kim, Yongsu and Kim, Howon},
  journal={IEEE Access}, 
  title={Cityscape-Adverse: Benchmarking Robustness of Semantic Segmentation with Realistic Scene Modifications via Diffusion-Based Image Editing}, 
  year={2025},
  doi={10.1109/ACCESS.2025.3537981}
}

ArXiv: arxiv.org/abs/2411.00425

Dataset Summary

• Base dataset: Cityscapes (finely annotated images; 2975 train / 500 val publicly available)
• Modifications:
  • Weathers: rainy, foggy
  • Seasons: spring, autumn, snowy
  • Lightings: sunny, night, dawn
• Generation:
  • Diffusion‑based instruction editing with CosXLEdit
  • Prompts e.g. "change the weather to foggy, photo-realistic", guidance scales 5–7, 20 diffusion steps.
• Resolution: 2048 × 1024 (same as Cityscapes)
• Labels: original Cityscapes semantic masks, preserved for all splits

Data Splits

Split	Source images	Modifications	Total images	Validation filters (accepted)	Acceptance rate
Train	2975 (Cityscapes train)	all 8 mods	2975 × 8 = 23 800	unfiltered	n/a
Validation	500 (Cityscapes val)	all 8 mods	4000	see below	93–100 %

Validation acceptance (after human filtering for credibility & semantic consistency):

• Rainy: 495 (99.0 %)
• Foggy: 478 (95.6 %)
• Spring: 493 (98.6 %)
• Autumn: 466 (93.2 %)
• Snowy: 488 (97.6 %)
• Sunny: 500 (100 %)
• Night: 500 (100 %)
• Dawn: 500 (100 %)

Total accepted: 3920 images.

Supported Tasks

• Semantic segmentation: evaluation of model robustness via mIoU, performance drop, and cross‑dataset generalization.

Licensing

• Images & labels: governed by the original Cityscapes license.
• Synthetic modifications: CC-BY-4.0
• Code & scripts: Apache 2.0

Acknowledgments

Supported by the MSIT, Korea, under the Convergence Security Core Talent Training Business (RS‑2022‑II221201), supervised by the IITP.

---

For questions or contributions, please open an issue or pull request on the GitHub repository.