v3 weights (yolo11s, 204k merged plate dataset, val mAP50=0.944)

miner.py

@@ -359,6 +359,51 @@ class Miner:
             dets.append((xa, ya, xb, yb, float(confs[i]), int(cls_ids[i])))
         return dets
 
+    def _cluster_dedup(
+        self,
+        dets: list[tuple[float, float, float, float, float, int]],
+        iou_thresh: float = 0.5,
+    ) -> list[tuple[float, float, float, float, float, int]]:
+        """Greedy near-duplicate suppression — for any pair with IoU >=
+        ``iou_thresh``, keep only the higher-conf detection.
+
+        Purpose: collapse TTA-induced duplicates of the same plate before
+        Soft-NMS, which would otherwise decay (but not kill) the lower-conf
+        copy, leaving multiple boxes per plate past our low score_threshold.
+        Mirrors the TTA-cluster-merge step in alfred8995/arabic000's miner.py.
+
+        Applied on *every* call (not just TTA) because the quad-4 overlap
+        band can also produce near-duplicate detections near tile seams.
+        IoU threshold 0.5 is loose enough that adjacent-but-distinct plates
+        (IoU < 0.5) stay separate; tight enough that same-plate variants
+        (IoU > 0.9 in practice) collapse.
+        """
+        if not dets:
+            return []
+        # Sort by conf desc (index 4)
+        srt = sorted(dets, key=lambda d: -d[4])
+        kept: list[tuple[float, float, float, float, float, int]] = []
+        suppressed = [False] * len(srt)
+        for i in range(len(srt)):
+            if suppressed[i]:
+                continue
+            x1i, y1i, x2i, y2i = srt[i][0], srt[i][1], srt[i][2], srt[i][3]
+            area_i = max(0.0, x2i - x1i) * max(0.0, y2i - y1i)
+            kept.append(srt[i])
+            for j in range(i + 1, len(srt)):
+                if suppressed[j]:
+                    continue
+                x1j, y1j, x2j, y2j = srt[j][0], srt[j][1], srt[j][2], srt[j][3]
+                ix1 = max(x1i, x1j); iy1 = max(y1i, y1j)
+                ix2 = min(x2i, x2j); iy2 = min(y2i, y2j)
+                iw = max(0.0, ix2 - ix1); ih = max(0.0, iy2 - iy1)
+                inter = iw * ih
+                area_j = max(0.0, x2j - x1j) * max(0.0, y2j - y1j)
+                union = area_i + area_j - inter
+                if union > 0 and inter / union >= iou_thresh:
+                    suppressed[j] = True
+        return kept
+
     def _quad4_raw_dets(
         self,
         image_bgr: ndarray,
@@ -415,6 +460,13 @@ class Miner:
                     (orig_w - x2f, y1, orig_w - x1f, y2, conf, cls_id)
                 )
 
+        # TTA-aware cluster-dedup: collapse near-duplicate detections of the
+        # same plate (e.g. original + unflipped TTA view) BEFORE Soft-NMS,
+        # which would otherwise decay but not kill the lower-conf copy at
+        # our low score_threshold=0.01. Without this step the deployed miner
+        # emitted 2-3 outputs per plate (verified on validator task 57820).
+        all_dets = self._cluster_dedup(all_dets, iou_thresh=0.5)
+
         dets = self._soft_nms(all_dets)
 
         out_boxes: list[BoundingBox] = []