vhr-buildings-automatic-annotation

18,557 image–mask pairs for training building segmentation models: 4,301 chips whose labels a human inspected and kept, plus 14,256 chips a label-quality classifier judged usable without human review.

A model trained on this pool reaches PQ 47.79 ± 0.23 on a held-out, human-reviewed benchmark, against 37.65 ± 1.29 for the same architecture trained on the human-reviewed chips alone — so the automatically filtered majority carries real signal rather than noise.

How the filtering works

Manual review established that 27.2% of the source dataset's labels are unusable (n = 6,459). Reviewing the rest by hand was not practical, so a classifier was trained on those decisions using nine measures of disagreement between a segmentation model and the label it was given:

IoU, precision, recall, label coverage, predicted coverage, mean image gradient along the label boundary, label-mask fragmentation, pixel variance inside the label, and the ratio of predicted to labelled connected components.

AUC 0.861 ± 0.011 under five-fold cross validation. Applied to the reviewed chips whose decision is known, the observed unusable rate is 43.9% and the mean predicted probability 45.9%, so the classifier is calibrated.

Image embeddings were tested against this and lost: DINOv2 and a task-trained encoder both reached about 0.77, below five columns of free metadata. Whether a label is wrong is a property of the label, and an image-only feature cannot see it.

Contents

source	chips	note
reviewed	4,301	human-inspected and kept
unreviewed	14,256	passed by the classifier at p(keep) >= 0.95

The threshold is strict on purpose: of 53,202 unreviewed chips, 26.8% passed at 0.95, against 69.0% at 0.5. Quote the threshold with any pass rate — the same classifier on the same data gives either number.

Myanmar is 13,667 of 18,557 chips (73.6%), inherited from the source dataset. Account for this — it is the dominant property of the pool, and the source dataset's validation and test splits contain no Myanmar at all.

Known issue: num_buildings is 0 for every unreviewed row

The column was not populated when unreviewed chips were joined by dataset_row, so all 14,256 report zero buildings. The masks are correct — in a 300-chip sample, 100% contain building pixels, with a mean mask area of 12.6% and a mean of 11.4 connected components per chip, making them denser than the reviewed portion. Derive building counts from the masks; do not stratify on num_buildings.

column	meaning
tile_id	OpenAerialMap tile identifier
dataset_row	row index into the source dataset
image	RGB chip, 256×256
mask	binary building mask
source	reviewed or unreviewed
num_buildings	unreliable — see above
country, project_name	from the source dataset

Evaluation

The 400 chips of nilsho01/vhr-buildings-benchmark-v1 are excluded from this pool, so it can be used as a held-out test set directly.

Report panoptic quality rather than pixel F1 or IoU. Pixel metrics cannot distinguish a correctly separated building from several merged into one: a single blob covering a terrace of four scores pixel IoU 86 and PQ 0, while the same four separated but traced one pixel too wide score pixel IoU 86 and PQ 86.

Limitations

These labels have not been seen by a human. At the threshold used, roughly one in twenty of the automatically kept chips is expected to carry a label a reviewer would have rejected. Where label fidelity matters more than volume, use vhr-buildings-verified-v1.

The classifier inherits the segmentation model's blind spots. Its features are disagreement between a model and a label, so a label wrong in a way that model is also wrong about will pass.

Geographic skew. Filtering raised the share of Mexico, Peru and Uganda but did not reduce Myanmar's dominance. A model trained here over-predicts out of domain: on the source dataset's test split, 37.4% of chips containing no buildings receive a spurious detection.

Related

• nilsho01/vhr-buildings-verified-v1 — human-reviewed chips only
• nilsho01/vhr-buildings-benchmark-v1 — the held-out benchmark
• nilsho01/vhr-buildings-review-decisions — the decisions the classifier learned from

Licence and attribution

ODbL, inherited from OpenStreetMap and OpenAerialMap via hotosm/vhr-building-segmentation. No other source contributed imagery or labels.