ObjectMatchingMetrics

ObjectMatchingMetrics compares predicted and reference label images by building pairwise object IoUs and greedily taking the best non-overlapping matches above iou_threshold.

Inputs are predicted_label_image, reference_label_image, and iou_threshold. Outputs report predicted/reference counts, matched and unmatched counts, and mean matched IoU/Dice.

Use it for deterministic instance-segmentation benchmark checks on small or medium label images.

Minimal Example

from bioimageflow_core import Arguments
from bioimageflow_measurement_tools import ObjectMatchingMetrics

metrics = ObjectMatchingMetrics().process_row(
    Arguments(
        predicted_label_image="predicted.tif",
        reference_label_image="reference.tif",
        iou_threshold=0.5,
    )
)

Inputs

• predicted_label_image: predicted label image.

• reference_label_image: reference label image with the same shape.

• iou_threshold: minimum object IoU for a match.

Outputs

• scalar counts for predicted, reference, matched, and unmatched objects.

• mean matched IoU and Dice over accepted matches.

Dependencies and Core Libraries

imageio and NumPy for label masks and pairwise overlap calculations.

Assumptions

Labels use 0 as background. Greedy matching is deterministic and adequate for smoke tests and simple benchmarks.

Expected Results

Synthetic fixtures with overlapping labels produce exact match, false-positive, and false-negative counts.

Failure Modes

Shape mismatches, unreadable images, and invalid labels raise errors before metrics are returned.