README.md

metadata

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:

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:

@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},
}