Compute a confusion matrix from SAHI predictions + COCO ground truth

For each image, greedily match predicted boxes to ground-truth boxes by IoU (highest-scoring prediction matches first). Predictions that match a GT box with IoU >= iou_threshold contribute a (gt_class, pred_class) cell. Unmatched predictions land in the "background" row (false positives); unmatched GT boxes land in the "background" column (false negatives).

Usage

confusion_matrix(
  predictions,
  annotation_json,
  iou_threshold = 0.5,
  score_threshold = 0
)

Arguments

predictions

Tibble of predictions with image_id, category_id, score, xmin, ymin, width, height (the predictions element of evaluate_model_sahi() output works directly).

annotation_json

Path to the COCO annotation JSON the predictions were produced against.

iou_threshold

Minimum IoU for a prediction to count as matching a GT box (default 0.5).

score_threshold

Filter predictions below this score before matching (default 0; keep all).

Value

A list with:

• long: long-format tibble (gt_label, pred_label, count)

• matrix: wide tibble, one row per GT label

• class_names: character vector of class labels (includes "background")

• iou_threshold: the threshold used

Details

Detection doesn't have a clean confusion-matrix definition (unmatched predictions and unmatched GT don't map onto classification confusion cleanly), so treat this as a diagnostic heatmap rather than a calibrated metric.