CoolFace
Datasetpublic

khalilurrahmanridoykhan/spacenet-buildings-vegas-smoketest-sample

SpaceNet Buildings (Vegas) — 20-Chip Smoke-Test Sample This is a tiny 20-chip sample, not a full dataset. It exists to reproducibly smoke-test a U-Net vs YOLO building-detection training pipeline end-to-end — not to train a real model. See the companion model repo khalilurrahmanridoykhan/unet-vs-yolo-buildings-smoketest and the source code at github.com/khalilurrrahmanridoykhan/best-model-for-satellite-imagery. Source and license Derived from SpaceNet Buildings v2… See the full description on the dataset page: https://huggingface.co/datasets/khalilurrahmanridoykhan/spacenet-buildings-vegas-smoketest-sample.

sourceHugging Facecc-by-sa-4.0updated 11d agoView on Hugging Face
0likes444downloads
Dataset Card

SpaceNet Buildings (Vegas) — 20-Chip Smoke-Test Sample

This is a tiny 20-chip sample, not a full dataset. It exists to reproducibly smoke-test a U-Net vs YOLO building-detection training pipeline end-to-end — not to train a real model. See the companion model repo `khalilurrahmanridoykhan/unet-vs-yolo-buildings-smoketest` and the source code at github.com/khalilurrrahmanridoykhan/best-model-for-satellite-imagery.

Source and license

Derived from SpaceNet Buildings v2 (AOI2Vegas), released under CC BY-SA 4.0 by the SpaceNet partners (Maxar/DigitalGlobe imagery, CosmiQ Works, and other SpaceNet contributors — see spacenet.ai for full attribution). This sample (20 chips, pulled from the public spacenet-dataset S3 bucket) and any derivatives are shared under the same CC BY-SA 4.0 license, per its share-alike terms.

The full dataset (thousands of chips) is not included here — pull it directly with `scripts/download_data.py` in the source repo (anonymous S3 access, no AWS account needed).

Contents

  • —raw/images/*.tif — 20 raw satellite chips (RGB, georeferenced)
  • —raw/labels/*.geojson — building-footprint polygons for each chip
  • —split.json — the train/val/test split (14/3/3) used for the smoke test, by whole chip (no chip appears in more than one split)

Rasterized U-Net masks and YOLO-format boxes are derived from raw/ by `notebooks/02_build_labels.ipynb` in the source repo rather than duplicated here.

Known data quirks (documented, not hidden)

A few of the underlying SpaceNet labels are degenerate Point or MultiPolygon geometries rather than clean Polygons — handled in the source pipeline's filter_polygon_geometries(), not filtered out of this raw sample.