WeatherSynthetic Driving Scene Dataset

WeatherSynthetic is a synthetic dataset featuring rich intrinsic map annotations, specifically designed for autonomous driving scenarios under diverse weather and lighting conditions. This dataset was introduced in our paper: "IntrinsicWeather: Controllable Weather Editing in Intrinsic Space". We hope it proves beneficial for future research in the field.

Dataset Overview

Metric	Value
Total Entries	35,035[^1]
Image Format	EXR (High Dynamic Range)
Annotation	Image + Intrinsic maps + Natural language description (Prompt)

[^1]:The paper mentions 38k entries; however, the public release contains 35,035 entries as certain scenes could not be released due to privacy constraints. We appreciate your understanding.

Directory Structure

The dataset contains 5 major scenes. Each scene has two subdirectories: image (HDR rendered images) and property (PBR material maps).

WeatherSynthetic/
├── Driving_prompts.json    # Main annotation file (image paths + text prompts)
├── Parking/                # Scene 1: Parking lot / underground garage
│   ├── image/
│   │   └── indoor/        # Indoor parking (varied lighting)
│   └── property/
│       ├── albedo/        # Base color maps (*.exr)
│       ├── metallic/      # Metallic maps (*.exr)
│       ├── normal/        # Normal maps (*.exr)
│       └── roughness/     # Roughness maps (*.exr)
├── Street/                # Scene 2: Street roads
│   ├── image/             # 8 weather conditions
│   └── property/          # albedo, metallic, normal, roughness
├── Town/                   # Scene 3: Town streets
│   ├── image/             # 9 weather conditions
│   └── property/          # albedo, metallic, normal, roughness
├── Small_city/             # Scene 4: Small city / urban plaza
│   ├── image/             # 9 weather conditions
│   └── property/          # albedo, metallic, normal, roughness
├── Modern_city/            # Scene 5: Modern urban streets
│   ├── image/             # 9 weather conditions
│   └── property/          # albedo, metallic, normal, roughness
└── README.md

Property Maps (PBR)

Each scene provides physically-based rendering (PBR) material maps aligned with the rendered images:

Map	Description
albedo	Base color / diffuse reflectance (e.g., 0000_albedo.exr)
metallic	Metallic workflow parameter
normal	Surface normal maps
roughness	Surface roughness maps

Property files share the same frame ID as images (e.g., 0000_image.exr ↔ 0000_albedo.exr).

Scene and Weather Types

Scene Types (5 Scenes)

Scene	Description	Image Layout	Weather/Lighting
Parking	Underground parking garage, indoor garage	image/indoor/	indoor
Street	Street roads	image/<weather>/	8 weather conditions
Town	Town streets	image/<weather>/	9 weather conditions
Small_city	Small city, urban plaza, autumn street views	image/<weather>/	9 weather conditions
Modern_city	Modern urban streets	image/<weather>/	9 weather conditions

Data Format

Driving_prompts.json

A JSON array where each element contains:

{
    "image_path": "WeatherSynthetic/Town/image/snowy/0000_image.exr",
    "prompt": "A vintage green streetcar glides through a snowy urban street on a cloudy winter afternoon."
}

Field	Type	Description
image_path	string	Relative path to the image, format: WeatherSynthetic/<scene>/image/<weather>/<id>_image.exr
prompt	string	English natural language description of the scene content, lighting, and weather

Path Conventions

• All paths are relative to the dataset root directory
• Image format: EXR (OpenEXR), HDR, suitable for lighting and weather research
• Property format: EXR. Property paths follow the same scene/frame structure; e.g., for Parking/image/indoor/0000_image.exr, the corresponding albedo is Parking/property/albedo/0000_albedo.exr

Usage Examples

We provide an example script to load, process, and visualize RGB image and intrinsic maps.

python -m data.WeatherSynthetic

Typical Applications

• Driving scene understanding under varied weather/lighting conditions
• Weather transfer and synthesis
• Robustness research for autonomous driving perception in adverse weather
• Text-guided scene editing and generation
• Physically-based rendering (PBR) and material editing (albedo, normal, etc.)
• Multimodal learning with HDR images and natural language

Dependencies

For reading EXR images:

• Python: OpenEXR or cv2 (OpenCV 4.x supports EXR)
• PyTorch: torchvision + cv2 or dedicated EXR libraries

License and Citation

Please comply with the relevant license terms when using this dataset. If used in academic work, please cite our paper in your publication.

@misc{zhu2026intrinsicweathercontrollableweatherediting,
      title={IntrinsicWeather: Controllable Weather Editing in Intrinsic Space}, 
      author={Yixin Zhu and Zuo-Liang Zhu and Jian Yang and Miloš Hašan and Jin Xie and Beibei Wang},
      year={2026},
      eprint={2508.06982},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.06982}, 
}