polymetrics

Fast object-level (polygon) accuracy metrics for geospatial ML. Compare two GeoJSONs and get precision, recall, F1, IoU, mAP, and polygon-shape statistics in one call.

uvx polymetrics truth.geojson pred.geojson

from polymetrics import evaluate

result = evaluate("truth.geojson", "pred.geojson", iou_threshold=0.5)
print(result.precision, result.recall, result.f1, result.mean_iou)
result.to_geojson("annotated.geojson")

What it computes

How it differs from pixel metrics

Pixel IoU pools every pixel and asks "what fraction did you get right?". Object IoU asks "did you find each building, and how well-shaped is each one?" which is what users actually see on a map. Both are useful for different reasons; this library only does object-level.

CRS handling

Inputs are always reprojected to EPSG:3857 before any geometric computation. If the prediction and truth GeoJSONs declare different source CRSs, a warning is emitted and both are projected. A missing CRS is treated as EPSG:4326 (the GeoJSON 2008 default).

Matching algorithm

Hungarian bipartite assignment (scipy.optimize.linear_sum_assignment) on a -IoU cost matrix, with gdf.sindex (STRtree) pruning to skip non-overlapping pairs. Each prediction matches at most one truth; pairs below iou_threshold are dropped (the prediction becomes FP, the truth FN).

mAP

mAP requires each prediction feature to carry a score or confidence property in [0, 1]. If neither is present, mAP is reported as null and a warning is emitted; the other metrics still compute.

Install

uv pip install polymetrics

CLI

polymetrics truth.geojson pred.geojson [options]

positional arguments:
  truth                 ground-truth GeoJSON
  pred                  prediction GeoJSON

options:
  --iou FLOAT           IoU threshold for TP (default: 0.5)
  --no-map              skip mAP computation
  --out PATH            write annotated GeoJSON (TP/FP/FN per feature)
  --json                emit metrics as JSON on stdout

License

Apache-2.0.