Fix AP/mAP calculations by sorting prediction confidence scores and handle IoU zero division (#545)

Co-authored-by: Shardul Mahamulkar <shardulmahamulkar@users.noreply.github.com>

Author: shardulmahamulkar

perceptionmetrics/utils/detection_metrics.py

@@ -127,9 +127,13 @@ def _match_predictions(
 
         ious = compute_iou_matrix(pred_boxes, gt_boxes)  # shape: (num_preds, num_gts)
 
-        for i, (p_box, p_label, score) in enumerate(
-            zip(pred_boxes, pred_labels, pred_scores)
-        ):
+        # Sort predictions by descending score to ensure highest-confidence predictions match first
+        sorted_indices = np.argsort([-s for s in pred_scores])
+
+        for i in sorted_indices:
+            p_box = pred_boxes[i]
+            p_label = pred_labels[i]
+            score = pred_scores[i]
             max_iou = 0
             max_j = -1