---
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>/`. The three image folders ship as zip archives so the repository stays within Hugging Face's file-count limits — `images/` is split into ~2 GB chunks. `splits/` and `LICENSE` are plain files.

```text
<Target_Domain>/
└── <Source_Domain>/
    ├── images_000.zip … images_NNN.zip   # Top 3 generated images per source (MCOC ranked)
    ├── original_tid.zip                  # Original source segmentation maps
    ├── hrda_tid.zip                      # HRDA pseudo-labels
    ├── splits/                           # .txt files for training loading
    └── LICENSE                           # Original source dataset license
```

Extracting the archives restores the layout the loading scripts expect. Entry names already carry their folder prefix (e.g. `images/03666/0005.png`), so every archive unpacks into the right place:

```text
<Target_Domain>/
└── <Source_Domain>/
    ├── images/
    ├── original_tid/
    ├── hrda_tid/
    ├── splits/
    └── LICENSE
```

| Pair | `images` chunks | Download |
| :--- | ---: | ---: |
| `cityscapes/gta` | 22 | 49.9 GB |
| `cityscapes/urbansyn` | 7 | 15.5 GB |
| `cityscapes/shift` | 3 | 6.1 GB |
| `cityscapes/veis` | 3 | 5.8 GB |
| `acdc/urbansyn` | 11 | 23.6 GB |
| `acdc/veis` | 4 | 9.0 GB |
| **Total** | **50** | **110 GB** |

---

## Download & Extraction

```bash
pip install -U "huggingface_hub[cli]"

# Everything (~110 GB)
hf download dkalsan/FLEDGE --repo-type dataset --local-dir FLEDGE

# Or a single pair, e.g. ACDC generated from VEIS (~9 GB)
hf download dkalsan/FLEDGE --repo-type dataset --local-dir FLEDGE --include "acdc/veis/*"
```

Then unpack every archive in place, deleting each one as it succeeds:

```bash
cd FLEDGE
for z in */*/*.zip; do
  unzip -q -o "$z" -d "$(dirname "$z")" && rm "$z"
done
```

---

## 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}
}
```