detection_accuracy#

geowombat.detect.detection_accuracy(predictions, truth, class_col='class_name', iou_thresholds=(0.5,), score_col='score', class_agnostic=False)[source]#

Compute detection accuracy metrics + a review-ready GeoDataFrame.

Parameters:
predictionsgeopandas.GeoDataFrame

Detector output. Must include geometry, score (or score_col), and a class column.

truthgeopandas.GeoDataFrame

Ground-truth boxes/polygons in the same CRS.

class_colstr

Column with class labels in both inputs.

iou_thresholdssequence of float, or ‘coco’

IoU thresholds to evaluate. 'coco' expands to 0.5..0.95 step 0.05 and returns mAP@[.5:.95].

score_colstr

Score column in predictions. Default ‘score’.

class_agnosticbool

If True, ignore class labels (treat as a single class).

Returns:
dict
Keys:

  • metrics: DataFrame indexed by class with columns precision, recall, f1, ap, tp, fp, fn, support for each IoU threshold.

  • summary: dict of overall mAP per threshold + mAP@[.5:.95].

  • matched: GeoDataFrame of TP/FP/FN rows ready for QGIS review (see export_for_review).

  • confusion: DataFrame of class confusion (per-class truth vs. matched-prediction class) at the lowest IoU.