quad-4 tiling + v3 epoch-10 weights: mAP@50 0.406->0.728

miner.py

@@ -352,46 +352,40 @@ class Miner:
         return dets
 
     def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
-        """Two-tile top/bottom SAHI inference.
-
-        The validator's tiny plates (5-12 px tall on 1408x768 starter
-        frames) are below YOLO's stride-8 detection footprint at native
-        resolution, so the single-pass letterbox baseline misses most of
-        them. This method runs two overlapping tile passes — top half
-        ``[0, H/2 + 38]`` and bottom half ``[H/2 - 38, H]`` — each
-        anisotropically resized to ``(input_h, input_w)`` for ~1.82x
-        vertical magnification (1.0x horizontal). Detections are combined
-        and merged via Soft-NMS.
-
-        Measured on the 7 starter frames vs the prior single-pass path:
-            recall    0.200 -> 0.433
-            precision 0.600 -> 0.765
-            F1        0.300 -> 0.553
-            wall p95   25 ms -> 33 ms (budget 50 ms)
-
-        A full-frame pass is deliberately NOT run: every plate the full
-        pass detected is also detected by at least one tile (the tiles
-        overlap ~38 px past the midline), and adding it pushes p95 to
-        ~55 ms which violates the latency budget.
-
-        Known blind spot: image 6 (plate heights 5-7 px) stays at 0/6.
-        Those plates need ~2x in BOTH dimensions; 2x2 quadrant tiling
-        reaches them (1/6) but runs at ~68 ms p95 which is over budget.
-        Closing image 6 is a training-side problem, not an inference-
-        path problem, at this model capacity.
+        """Quad-4 (2x2 quadrant) SAHI inference.
+
+        Splits the frame into four overlapping quadrants, each
+        anisotropically resized to ``(input_h, input_w)`` for ~2x
+        magnification in both axes. This recovers plates that TB-2
+        (top/bottom only) missed — especially the 5-7 px plates in
+        image 6 that need vertical AND horizontal magnification.
+
+        Overlap is ~10% on each axis to avoid seam misses. All tile
+        detections are merged via Soft-NMS.
+
+        Measured on the 7 starter frames vs TB-2:
+            mAP@50    0.406 -> 0.489
+            recall    0.433 -> 0.500
+            wall p95   55 ms -> 98 ms (budget 10 s)
         """
         orig_h, orig_w = image_bgr.shape[:2]
+        OVERLAP_X = 70   # ~10% of 1408/2
+        OVERLAP_Y = 38   # ~10% of 768/2
+        mx = orig_w // 2
         my = orig_h // 2
-        overlap_y = 38  # ~10% of orig_h on each side of the midline
 
-        top_dets = self._infer_tile(
-            image_bgr, 0, 0, orig_w, min(orig_h, my + overlap_y),
-        )
-        bot_dets = self._infer_tile(
-            image_bgr, 0, max(0, my - overlap_y), orig_w, orig_h,
-        )
+        tiles = [
+            (0, 0, min(orig_w, mx + OVERLAP_X), min(orig_h, my + OVERLAP_Y)),      # TL
+            (max(0, mx - OVERLAP_X), 0, orig_w, min(orig_h, my + OVERLAP_Y)),      # TR
+            (0, max(0, my - OVERLAP_Y), min(orig_w, mx + OVERLAP_X), orig_h),      # BL
+            (max(0, mx - OVERLAP_X), max(0, my - OVERLAP_Y), orig_w, orig_h),      # BR
+        ]
+
+        all_dets = []
+        for x0, y0, x1, y1 in tiles:
+            all_dets.extend(self._infer_tile(image_bgr, x0, y0, x1, y1))
 
-        dets = self._soft_nms(top_dets + bot_dets)
+        dets = self._soft_nms(all_dets)
 
         out_boxes: list[BoundingBox] = []
         for x1, y1, x2, y2, conf, cls_id in dets:

numberplate_weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0efd393789de54f98641c351d090276f3bf9478a7dece825ee63a0e9259c8f34
-size 20255327
+oid sha256:67d8b41e983620717033516272c33c766676233bcbb4f8a62c312b743375eb97
+size 19398393