YOLO26s + SAHI 2x2 tiling, c=0.20, shutdown=14400

miner.py

@@ -36,11 +36,6 @@ class TVFrameResult(BaseModel):
 
 
 class Miner:
-    """
-    Auto-generated by subnet_bridge from a Manako element repo.
-    This miner is intentionally self-contained for chute import restrictions.
-    """
-
     def __init__(self, path_hf_repo: Path) -> None:
         self.path_hf_repo = path_hf_repo
         self.class_names = ['numberplate']
@@ -50,99 +45,123 @@ class Miner:
         )
         self.input_name = self.session.get_inputs()[0].name
         input_shape = self.session.get_inputs()[0].shape
-        self.input_h = int(input_shape[2])
-        self.input_w = int(input_shape[3])
-        self.conf_threshold = 0.15
-        self.iou_threshold = 0.3
+        self.input_size = int(input_shape[2])  # 1280
+        self.conf_threshold = 0.20
+        self.iou_threshold = 0.5
+        self.tile_overlap = 0.2
 
     def __repr__(self) -> str:
         return f"ONNX Miner session={type(self.session).__name__} classes={len(self.class_names)}"
 
-    def _preprocess(self, image_bgr: ndarray) -> tuple[np.ndarray, tuple[int, int]]:
-        h, w = image_bgr.shape[:2]
-        rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
-        resized = cv2.resize(rgb, (self.input_w, self.input_h))
-        x = resized.astype(np.float32) / 255.0
-        x = np.transpose(x, (2, 0, 1))[None, ...]
-        return x, (h, w)
-
-    def _normalize_predictions(self, raw: np.ndarray) -> np.ndarray:
-        pred = raw[0]
-        if pred.ndim != 2:
-            raise ValueError(f"Unexpected prediction shape: {raw.shape}")
-        if pred.shape[0] < pred.shape[1]:
-            pred = pred.transpose(1, 0)
-        return pred
-
-    def _nms(self, dets: list[tuple[float, float, float, float, float, int]]) -> list[tuple[float, float, float, float, float, int]]:
+    def _letterbox(self, img: ndarray) -> tuple[ndarray, float, int, int]:
+        h, w = img.shape[:2]
+        r = min(self.input_size / h, self.input_size / w)
+        nw, nh = int(w * r), int(h * r)
+        resized = cv2.resize(img, (nw, nh), interpolation=cv2.INTER_LINEAR)
+        canvas = np.full((self.input_size, self.input_size, 3), 114, dtype=np.uint8)
+        dx = (self.input_size - nw) // 2
+        dy = (self.input_size - nh) // 2
+        canvas[dy:dy+nh, dx:dx+nw] = resized
+        return canvas, r, dx, dy
+
+    def _run_single(self, img: ndarray) -> list[tuple[float, float, float, float, float]]:
+        h, w = img.shape[:2]
+        canvas, r, dx, dy = self._letterbox(img)
+        blob = (canvas.astype(np.float32) / 255.0).transpose(2, 0, 1)[np.newaxis]
+        out = self.session.run(None, {self.input_name: blob})[0][0]
+        dets = []
+        for row in out:
+            x1, y1, x2, y2, conf, cls = row
+            if conf < self.conf_threshold:
+                continue
+            dets.append((
+                float(conf),
+                (x1 - dx) / r,
+                (y1 - dy) / r,
+                (x2 - dx) / r,
+                (y2 - dy) / r,
+            ))
+        return dets
+
+    def _nms(self, dets: list[tuple[float, float, float, float, float]]) -> list[tuple[float, float, float, float, float]]:
         if not dets:
             return []
-        boxes = np.array([[d[0], d[1], d[2], d[3]] for d in dets], dtype=np.float32)
-        scores = np.array([d[4] for d in dets], dtype=np.float32)
-        order = scores.argsort()[::-1]
+        dets.sort(key=lambda x: -x[0])
         keep = []
-        while order.size > 0:
-            i = order[0]
-            keep.append(i)
-            xx1 = np.maximum(boxes[i, 0], boxes[order[1:], 0])
-            yy1 = np.maximum(boxes[i, 1], boxes[order[1:], 1])
-            xx2 = np.minimum(boxes[i, 2], boxes[order[1:], 2])
-            yy2 = np.minimum(boxes[i, 3], boxes[order[1:], 3])
-            w = np.maximum(0.0, xx2 - xx1)
-            h = np.maximum(0.0, yy2 - yy1)
-            inter = w * h
-            area_i = (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
-            area_rest = (boxes[order[1:], 2] - boxes[order[1:], 0]) * (boxes[order[1:], 3] - boxes[order[1:], 1])
-            union = np.maximum(area_i + area_rest - inter, 1e-6)
-            iou = inter / union
-            remaining = np.where(iou <= self.iou_threshold)[0]
-            order = order[remaining + 1]
-        return [dets[idx] for idx in keep]
+        used = [False] * len(dets)
+        for i in range(len(dets)):
+            if used[i]:
+                continue
+            keep.append(dets[i])
+            for j in range(i + 1, len(dets)):
+                if used[j]:
+                    continue
+                # compute IoU
+                ax1, ay1, ax2, ay2 = dets[i][1], dets[i][2], dets[i][3], dets[i][4]
+                bx1, by1, bx2, by2 = dets[j][1], dets[j][2], dets[j][3], dets[j][4]
+                ix1 = max(ax1, bx1); iy1 = max(ay1, by1)
+                ix2 = min(ax2, bx2); iy2 = min(ay2, by2)
+                inter = max(0, ix2-ix1) * max(0, iy2-iy1)
+                aa = (ax2-ax1)*(ay2-ay1); bb = (bx2-bx1)*(by2-by1)
+                iou = inter / (aa + bb - inter + 1e-6)
+                if iou > self.iou_threshold:
+                    used[j] = True
+        return keep
 
     def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
-        inp, (orig_h, orig_w) = self._preprocess(image_bgr)
-        out = self.session.run(None, {self.input_name: inp})[0]
-        pred = self._normalize_predictions(out)
-        if pred.shape[1] < 5:
-            return []
-        boxes = pred[:, :4]
-        cls_scores = pred[:, 4:]
-        if cls_scores.shape[1] == 0:
-            return []
-        cls_ids = np.argmax(cls_scores, axis=1)
-        confs = np.max(cls_scores, axis=1)
-        keep = confs >= self.conf_threshold
-        boxes = boxes[keep]
-        confs = confs[keep]
-        cls_ids = cls_ids[keep]
-        if boxes.shape[0] == 0:
-            return []
-        sx = orig_w / float(self.input_w)
-        sy = orig_h / float(self.input_h)
-        dets: list[tuple[float, float, float, float, float, int]] = []
-        for i in range(boxes.shape[0]):
-            cx, cy, bw, bh = boxes[i].tolist()
-            x1 = (cx - bw / 2.0) * sx
-            y1 = (cy - bh / 2.0) * sy
-            x2 = (cx + bw / 2.0) * sx
-            y2 = (cy + bh / 2.0) * sy
-            dets.append((x1, y1, x2, y2, float(confs[i]), int(cls_ids[i])))
-        dets = self._nms(dets)
-        out_boxes: list[BoundingBox] = []
-        for x1, y1, x2, y2, conf, cls_id in dets:
-            ix1 = max(0, min(orig_w, math.floor(x1)))
-            iy1 = max(0, min(orig_h, math.floor(y1)))
-            ix2 = max(0, min(orig_w, math.ceil(x2)))
-            iy2 = max(0, min(orig_h, math.ceil(y2)))
+        orig_h, orig_w = image_bgr.shape[:2]
+        all_dets = []
+
+        # Full image pass
+        all_dets.extend(self._run_single(image_bgr))
+
+        # 2x2 tile passes
+        tw = orig_w // 2
+        th = orig_h // 2
+        ox = int(tw * self.tile_overlap)
+        oy = int(th * self.tile_overlap)
+        tiles = [
+            (0, 0, tw + ox, th + oy),
+            (tw - ox, 0, orig_w, th + oy),
+            (0, th - oy, tw + ox, orig_h),
+            (tw - ox, th - oy, orig_w, orig_h),
+        ]
+        for tx1, ty1, tx2, ty2 in tiles:
+            tx1 = max(0, tx1); ty1 = max(0, ty1)
+            tx2 = min(orig_w, tx2); ty2 = min(orig_h, ty2)
+            crop = image_bgr[ty1:ty2, tx1:tx2]
+            tile_dets = self._run_single(crop)
+            for conf, x1, y1, x2, y2 in tile_dets:
+                all_dets.append((conf, x1 + tx1, y1 + ty1, x2 + tx1, y2 + ty1))
+
+        # NMS to merge overlapping detections
+        all_dets = self._nms(all_dets)
+
+        out_boxes = []
+        for conf, x1, y1, x2, y2 in all_dets:
+            bx1 = max(0, min(orig_w, math.floor(x1)))
+            by1 = max(0, min(orig_h, math.floor(y1)))
+            bx2 = max(0, min(orig_w, math.ceil(x2)))
+            by2 = max(0, min(orig_h, math.ceil(y2)))
+            bw = bx2 - bx1
+            bh = by2 - by1
+            if bw < 6 or bh < 6 or bw * bh < 80:
+                continue
+            if max(bw / max(bh, 1), bh / max(bw, 1)) > 10:
+                continue
             out_boxes.append(
-                BoundingBox(x1=ix1, y1=iy1, x2=ix2, y2=iy2,
-                    cls_id=cls_id, conf=max(0.0, min(1.0, conf))))
+                BoundingBox(
+                    x1=bx1, y1=by1, x2=bx2, y2=by2,
+                    cls_id=0,
+                    conf=max(0.0, min(1.0, conf)),
+                )
+            )
         return out_boxes
 
     def predict_batch(
         self, batch_images: list[ndarray], offset: int, n_keypoints: int,
     ) -> list[TVFrameResult]:
-        results: list[TVFrameResult] = []
+        results = []
         for idx, image in enumerate(batch_images):
             boxes = self._infer_single(image)
             keypoints = [(0, 0) for _ in range(max(0, int(n_keypoints)))]

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:97d257baf273fe85d6b51ea547b59dd36d2a67a60fa64e30c9c15ce7d7949995
-size 11966914
+oid sha256:16b083d01cf9fff412433a8d3bb1bd9253aa29cefabb05a2afc83c5fe55c520d
+size 19405465