NeurIPS2026EDTrack/PolyTopoBench

PolyTopoBench

PolyTopoBench is a remote-sensing benchmark for evaluating vector polygon generation with both exterior boundaries and interior rings. The dataset release contains two aligned sources used in the benchmark:

• —inria_dataset_aligned: aligned aerial imagery, building masks, and polygon annotations derived from the Inria building dataset.
• —deventer_512_valtest_as_val: 512 x 512 aerial image tiles with polygon annotations for multiple land-cover classes; validation and test splits are merged as the validation split used by the benchmark.

Files

• —inria_dataset_aligned/ and deventer_512_valtest_as_val/: full released datasets.
• —SampleData/: representative subset for quick inspection by reviewers. The sample keeps the same folder structure as the full data and was created by selecting a small number of image tiles and their corresponding masks/annotations from each dataset.
• —metadata/: Croissant metadata files with core and Responsible AI fields.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="NeurIPS2026EDTrack/PolyTopoBench",
    repo_type="dataset",
    local_dir="dataset"
)

The released code expects the downloaded folders under dataset/. After downloading, run the repository prepare_data.py script to convert the raw folders into the processed format used by the benchmark baselines and evaluator.

Code

The benchmark code is available at: https://anonymous.4open.science/r/PolyTopoBench-Anonymized