Cityscape-Adverse: Semantic Segmentation Benchmark Under Eight Realistic Conditions

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. The dataset was created by author 'naufalso' and was last updated on Hugging Face on May 14, 2025.

Use Cases

• Benchmark model robustness to adverse weather based on rainy, foggy, and snowy scene modifications
• Evaluate segmentation performance across different times of day based on night and dawn scene modifications
• Test model generalization across seasonal variations based on spring and autumn scene modifications
• Assess performance under varying lighting conditions based on sunny scene modifications

Strengths

• Applies eight distinct, realistic environmental modifications to a standard benchmark dataset
• Preserves original high-resolution 2048×1024 semantic labels for direct evaluation
• Uses diffusion-based image editing for realistic transformations

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download
• Row count, file formats, and column-level documentation are absent

Provenance

Source

naufalso on Hugging Face

Collection Method

Extends the original Cityscapes dataset using diffusion-based image editing.

Freshness

Last updated 2025-05-14 02:21:36; freshness should be verified

Geography

Urban scenes, likely from the original Cityscapes dataset