chrise222/WeatherSynthetic

---
license: cc-by-nc-4.0
task_categories:
- text-to-image
- image-to-image
- image-text-to-image
size_categories:
- 10K<n<100K
viewer: false
---
# 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:

```json
{
"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 an RGB image and intrinsic maps.
Please following the instruction in [code](https://github.com/YixinZhu042/IntrinsicWeather).

```bash
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.

```bibtex
@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},
}
```

---
许可证：CC BY-NC 4.0
任务类别：
- 文本到图像
- 图像到图像
- 图像-文本到图像
规模类别：
- 10K<n<100K
查看器：已禁用
---
# WeatherSynthetic 驾驶场景合成数据集

WeatherSynthetic是一款富含本征映射（intrinsic map）标注的合成数据集，专为多样化天气与光照条件下的自动驾驶（autonomous driving）场景设计。本数据集随我们的论文《IntrinsicWeather：本征空间中的可控天气编辑》一同发布。我们期望该数据集能够为该领域的后续研究提供助力。

## 数据集概览

| 指标 | 数值 |
|--------|-------|
| 总条目数 | 35,035[^1] |
| 图像格式 | EXR（OpenEXR，高动态范围（High Dynamic Range, HDR）） |
| 标注内容 | 图像 + 本征映射 + 自然语言描述（提示词Prompt） |

[^1]：论文中提及该数据集共有3.8万条条目，但由于部分场景涉及隐私限制无法公开，本次公开发布的版本实际包含35,035条条目，敬请谅解。

## 目录结构

本数据集包含**5大场景**，每个场景下设两个子目录：**image**（高动态范围渲染图像）与**property**（基于物理的渲染（Physically-based Rendering, PBR）材质映射）。

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

### 基于物理的渲染（Physically-based Rendering, PBR）材质映射

每个场景均提供与渲染图像对齐的基于物理的渲染材质映射：

| 映射类型 | 描述 |
|-----|-------------|
| **反照率（albedo）** | 基础颜色/漫反射率（例如`0000_albedo.exr`） |
| **金属度（metallic）** | 金属度工作流参数 |
| **法向映射（normal）** | 表面法向映射 |
| **粗糙度（roughness）** | 表面粗糙度映射 |

属性文件与图像文件使用相同的帧ID（例如`0000_image.exr` ↔ `0000_albedo.exr`）。

## 场景与天气类型

### 场景类型（共5类场景）

| 场景名称 | 场景描述 | 图像路径结构 | 天气/光照条件 |
|-------|-------------|--------------|------------------|
| **停车场（Parking）** | 地下停车场、室内车库 | `image/indoor/` | 室内光照 |
| **城市道路（Street）** | 城市街道 | `image/<weather>/` | 8种天气条件 |
| **城镇街道（Town）** | 城镇街道 | `image/<weather>/` | 9种天气条件 |
| **小型城市（Small_city）** | 小型城市、城市广场、秋日街景 | `image/<weather>/` | 9种天气条件 |
| **现代城市（Modern_city）** | 现代城市街道 | `image/<weather>/` | 9种天气条件 |

## 数据格式

### Driving_prompts.json

该文件为JSON数组，其中每个元素包含以下字段：

json
{
"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."
}

| 字段名 | 数据类型 | 描述 |
|-------|------|--------------|
| `image_path` | 字符串 | 图像相对路径，格式为`WeatherSynthetic/<scene>/image/<weather>/<id>_image.exr` |
| `prompt` | 字符串 | 描述场景内容、光照与天气的英文自然语言文本 |

### 路径约定

- 所有路径均相对于数据集根目录
- **图像格式**：EXR（OpenEXR）格式，高动态范围（HDR），适用于光照与天气相关研究
- **属性文件格式**：EXR格式。属性文件的路径遵循与图像文件一致的场景/帧结构；例如，针对`Parking/image/indoor/0000_image.exr`，其对应的反照率映射文件路径为`Parking/property/albedo/0000_albedo.exr`

## 使用示例

我们提供了示例脚本，用于加载、处理并可视化RGB图像与本征映射。请遵循[代码仓库](https://github.com/YixinZhu042/IntrinsicWeather)中的说明进行操作。

bash
python -m data.WeatherSynthetic

## 典型应用场景

- 多样化天气/光照条件下的驾驶场景理解
- 天气迁移与合成
- 自动驾驶感知模型在恶劣天气下的鲁棒性研究
- 文本引导的场景编辑与生成
- 基于物理的渲染（PBR）与材质编辑（反照率、法向映射等）
- 结合高动态范围图像与自然语言的多模态学习

## 依赖库

用于读取EXR图像的依赖库：
- **Python**：`OpenEXR`库或`cv2`（OpenCV 4.x及以上版本支持EXR格式）
- **PyTorch**：`torchvision` + `cv2`，或专用的EXR处理库

## 许可证与引用

使用本数据集时，请遵守相关许可证条款。若将本数据集用于学术研究，请在发表成果中引用我们的论文。

bibtex
@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},
}