dkalsan/FLEDGE
收藏Hugging Face2026-04-20 更新2026-04-26 收录
下载链接：
https://hf-mirror.com/datasets/dkalsan/FLEDGE
下载链接
链接失效反馈官方服务：
资源简介：
---
license: gpl-3.0
task_categories:
- image-segmentation
tags:
- semantic-segmentation
- synthetic-data
- domain-adaptation
- cityscapes
- acdc
size_categories:
- 100K<n<1M
---
## Dataset Summary
This dataset contains generated images for semantic segmentation targeting the **Cityscapes** and **ACDC** domains. The images are produced from five source datasets using the method described in: [A Framework for Low-Effort Training Data Generation for Urban Semantic Segmentation](https://arxiv.org/abs/2510.11567).
## Technical Specifications
* **Subset Selection:** To preserve storage space, we release the **top 3 generated images** per original source image, selected based on our proposed **MCOC score**.
* **Label Mapping:** All masks follow the standard **Cityscapes 19-class format**.
* **Resolution:** 512 x 1024 (Height x Width).
* **Format:** Images and masks are provided in `.png` format.
---
## Dataset Statistics
The following table reflects the number of source images used to generate the target domain data.
| Source Domain | Image Count | Status |
| :--- | :--- | :--- |
| **GTA** | 24,966 | Available |
| **UrbanSyn** | 7,539 | Available |
| **VEIS** | 3,018 | Available |
| **SHIFT** | 3,000 | Available |
| **Synscapes** | 0 | Restricted |
> **Note on Synscapes:** Due to licensing restrictions, we are unable to release the generated Synscapes data. The folders contain the original license file for reference but do not include image data.
---
## Dataset Structure
The data is organized by `<Target_Domain>/<Source_Domain>/`.
```text
<Target_Domain>/
└── <Source_Domain>/
├── images/ # Top 3 generated images per source (MCOC ranked)
├── original_tid/ # Original source segmentation maps
├── hrda_tid/ # HRDA pseudo-labels
├── splits/ # .txt files for training loading
└── LICENSE # Original source dataset license
```
---
## Usage Example
Example data loading scripts that utilize the provided .txt split files are available in our official GitHub repository: [github.com/vislearn/FLEDGE](https://github.com/vislearn/FLEDGE)
---
## Citation
```text
@misc{kalsan2025fledge,
title={A Framework for Low-Effort Training Data Generation for Urban Semantic Segmentation},
author={Damjan Kalšan and Denis Zavadski and Tim Küchler and Haebom Lee and Stefan Roth and Carsten Rother},
year={2025},
eprint={2510.11567},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
提供机构：
dkalsan
