HeinzJiao/Deventer-512

---
pretty_name: Deventer-512
language:
- en
license: cc-by-nc-4.0
task_categories:
- image-segmentation
- object-detection
tags:
- remote-sensing
- aerial-imagery
- orthophoto
- polygon-extraction
- polygonal-vectorization
- all-class-polygonal-vectorization
- acpv
- topology
- semantic-segmentation
- benchmark
size_categories:
- 1K<n<10K
---

# Deventer-512

## Dataset Summary

Deventer-512 is the benchmark dataset introduced in our paper *ACPV-Net: All-Class Polygonal Vectorization for Seamless Vector Map Generation from Aerial Imagery*. It is the first public benchmark for **All-Class Polygonal Vectorization (ACPV)**, a task that aims to generate a complete vector map from aerial imagery in a single run by producing polygons for all land-cover classes with **shared boundaries** and **without gaps or overlaps**.

The dataset contains **2,148** orthophoto tiles of size **512 x 512** pixels, together with raster masks and per-class COCO-style polygon annotations. It is designed for standardized evaluation of semantic fidelity, geometric accuracy, vertex efficiency, per-class topological fidelity, and global topological consistency.

The benchmark is organized around five semantically meaningful urban land-cover categories:

- `building`
- `road`
- `vegetation`
- `water`
- `unvegetated`

## Supported Tasks

Deventer-512 supports the following research tasks:

- **All-Class Polygonal Vectorization (ACPV)**: seamless multi-class vector map generation over the full image domain with shared boundaries and no gaps or overlaps
- **Multi-class semantic segmentation**: dense semantic prediction using the provided raster masks
- **Single-class polygonal vectorization**: category-wise polygon extraction such as building outline extraction or road region vectorization
- **Instance segmentation and object detection**: supported for categories and settings where COCO-style polygon annotations are appropriate

## Data Composition

Each split follows the same directory structure:

```text
deventer_512/
├── train/
│ ├── images/
│ ├── masks/
│ └── annotations/
├── val/
│ ├── images/
│ ├── masks/
│ └── annotations/
└── test/
├── images/
├── masks/
└── annotations/
```

### Splits

The official split sizes are:

| Split | Number of tiles |
|-------|-----------------:|
| train | 1716 |
| val | 212 |
| test | 220 |

Total: **2,148** image tiles.

### Files in Each Split

- `images/`: RGB orthophoto tiles in PNG format
- `masks/`: raster semantic masks aligned with the image tiles
- `annotations/`: per-class COCO-style polygon annotations

The `annotations/` folder contains one JSON file per class:

- `building.json`
- `road.json`
- `vegetation.json`
- `water.json`
- `unvegetated.json`

## Annotation Format

Polygon annotations are stored in standard COCO-style JSON format. Each annotation file corresponds to a single semantic category and contains:

- `categories`
- `images`
- `annotations`

Each `images` entry provides:

- `id`
- `file_name`
- `height`
- `width`

Each `annotations` entry provides:

- `id`
- `image_id`
- `category_id`
- `segmentation`
- `area`
- `bbox`
- `iscrowd`

## Citation

If you use Deventer-512 in your research, please cite:

```bibtex
@misc{jiao2026acpvnetallclasspolygonalvectorization,
title={ACPV-Net: All-Class Polygonal Vectorization for Seamless Vector Map Generation from Aerial Imagery},
author={Weiqin Jiao and Hao Cheng and George Vosselman and Claudio Persello},
year={2026},
eprint={2603.16616},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.16616},
}
```