Upload folder using huggingface_hub

miner.py

@@ -24,141 +24,254 @@ class TVFrameResult(BaseModel):
 
 
 class Miner:
-    """
-    YOLOv26 ONNX miner for car wash detection.
-    
-    Classes: broom, drainage gate, nozzle, track
-    
-    v26 is NMS-free — output shape: [1, 300, 6] (x1, y1, x2, y2, conf, cls_id).
-    
-    Features:
-      - Vectorized NMS + sanity filter + dedup + flip TTA
-      - Per-class rescue bonus (saves hard-to-detect classes at slightly lower conf)
-      - Confidence boost from same-class cluster (TTA consensus)
-      - Aggressive same-class overlap suppression
-      - Per-class IoU thresholds
-    """
-
-    class_names = ['broom', 'drainage gate', 'nozzle', 'track']
-    input_size = 1408
-    cross_iou_thresh = 0.8
-    max_det = 300
-    #overlap_suppress_threshold = 0.85
-    
-    # Per-class confidence thresholds
-    _conf_thres_array = np.array([0.35, 0.7, 0.4, 0.7], dtype=np.float32)
-    _extra_conf_thres_array = np.array([0.35, 0.35, 0.55, 0.35], dtype=np.float32)
-    
-    # Per-class IoU thresholds for same-class NMS
-    _iou_thres_array = np.array([0.6, 0.65, 0.5, 0.65], dtype=np.float32)
-    
-    # Per-class rescue bonus
-    _bonus_array = np.array([0.2, 0.2, 0.0, 0.2], dtype=np.float32)
-    
-    # Per-class minimum box area
-    # Indices: 0=broom, 1=drainage gate, 2=nozzle, 3=track
-    _min_box_area_array = np.array([144.0, 144.0, 4.0, 144.0], dtype=np.float32)
-    
-
-    def __init__(self, path_hf_repo: Path) -> None:
-        self.path_hf_repo = path_hf_repo
-        
+    def __init__(self,
+        path_hf_repo: Path
+    ) -> None:
+        model_path = self._resolve_model_path(path_hf_repo)
+        # car-wash element classes — cls_id order MUST match element `objects`
+        # (0=broom, 1=drainage gate, 2=nozzle, 3=track). This is the canonical
+        # order every downstream consumer (validator, BoundingBox.cls_id) sees.
+        self.class_names = ["broom", "drainage gate", "nozzle", "track"]
+        # FALLBACK model-emit order: the authoritative order is read from the
+        # ONNX `names` metadata after the session is created (embedded by
+        # Ultralytics at export, ships inside weights.onnx), so a retrained
+        # model with a different class order is remapped correctly without
+        # code changes. This list is used only when metadata is missing.
+        self._model_class_order = ["broom", "drainage gate", "nozzle", "track"]
         print("ORT version:", ort.__version__)
-        
+
         try:
             ort.preload_dlls()
-            print("preload_dlls success")
+            print("✅ onnxruntime.preload_dlls() success")
         except Exception as e:
-            print(f"preload_dlls failed: {e}")
-        
+            print(f"⚠️ preload_dlls failed: {e}")
+
         print("ORT available providers BEFORE session:", ort.get_available_providers())
-        
+
         sess_options = ort.SessionOptions()
         sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-        
-        self.session = ort.InferenceSession(
-            str(path_hf_repo / "weights.onnx"),
-            sess_options=sess_options,
-            providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
-        )
-        print("Created ORT session with preferred CUDA provider list")
+        sess_options.intra_op_num_threads = 2
+        sess_options.inter_op_num_threads = 1
+        sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+
+        try:
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+            )
+            print("✅ Created ORT session with preferred CUDA provider list")
+        except Exception as e:
+            print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}")
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )
+
         print("ORT session providers:", self.session.get_providers())
-        
+
+        # Build cls_remap: for each model-emit index i,
+        #   cls_remap[i] = self.class_names.index(model_class_order[i])
+        # The model-side order comes from the ONNX metadata when available,
+        # else falls back to the static _model_class_order.
+        model_class_order = self._read_model_class_order()
+        if model_class_order is None:
+            model_class_order = list(self._model_class_order)
+            print(f"cls order: no usable ONNX metadata, FALLBACK {model_class_order}")
+        else:
+            print(f"cls order: from ONNX metadata {model_class_order}")
+        self.cls_remap = np.array(
+            [self.class_names.index(n) for n in model_class_order], dtype=np.int32
+        )
+
+        for inp in self.session.get_inputs():
+            print("INPUT:", inp.name, inp.shape, inp.type)
+
+        for out in self.session.get_outputs():
+            print("OUTPUT:", out.name, out.shape, out.type)
+
         self.input_name = self.session.get_inputs()[0].name
         self.output_names = [output.name for output in self.session.get_outputs()]
-        input_shape = self.session.get_inputs()[0].shape
-        
-        self.input_h = self._safe_dim(input_shape[2], default=self.input_size)
-        self.input_w = self._safe_dim(input_shape[3], default=self.input_size)
+        self.input_shape = self.session.get_inputs()[0].shape
+
+        # Match the ONNX input dtype (this export is FP16 -> needs float16 input).
+        input_type = self.session.get_inputs()[0].type
+        self.np_dtype = np.float16 if "float16" in input_type else np.float32
+        print(f"✅ ONNX input dtype: {input_type} -> numpy {self.np_dtype}")
+
+        # ONNX is fixed-size 1408x1408 (v1 export); read actual shape to be safe.
+        self.input_height = self._safe_dim(self.input_shape[2], default=1280)
+        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
+
+        # Tuned for validator scoring (pillars: 0.6*map50 + 0.4*false_positive).
+        # All values below are the measured optimum of a full grid sweep on
+        # the validator-style val split (tune_miner.py, 241 1024x1024 crops,
+        # composite 0.8002 -> 0.8103) -- re-run the sweep after any retrain.
+        self.iou_thres = 0.5        # Per-class NMS IoU; lower = stricter dedup
+        self.cross_iou_thresh = 0.9  # Cross-class dedup IoU (suppress same physical object firing multiple classes)
+        self.max_det = 200
+        # TTA = a 2nd (flipped) forward pass. Doubles latency; off for the
+        # CPU latency gate. Re-enable only if the latency budget allows.
+        self.use_tta = True
+
+#       conf thresholds:  broom=0.38  drainage gate=0.45  nozzle=0.30  track=0.60
+        # Per-class confidence thresholds.
+        # Indexed by class_names order: [broom, drainage gate, nozzle, track].
+        # broom/nozzle sit low: under the validator metric the mAP gained
+        # from the extra recall outweighs the FP-pillar cost (the previous
+        # 0.5/0.5 silently discarded many valid detections); track is the
+        # one class where false fires are common enough to need 0.38.
+        self._conf_thres_array = np.array(
+            [0.28, 0.38, 0.60, 0.45], dtype=np.float32
+        )
+        # Per-class rescue bonus: when a class has ZERO boxes passing the
+        # threshold in a frame, its top-1 candidate is admitted when its score
+        # is at least (per-class threshold - per-class bonus).
+        # DISABLED (all zeros): the sweep showed rescue admits more false
+        # positives than true positives under the validator's FP pillar.
+        self._bonus_array = np.array(
+            [0.05, 0.1, 0.25, 0.2], dtype=np.float32
+        )
+
+        # Box sanity filter — kept loose: car-wash `nozzle` boxes are tiny
+        # (GT median ~290 px², smallest ~32 px²). Fire's 14x14/min_side 8
+        # would delete valid nozzles, so thresholds are dropped here.
+        self.min_box_area = 4 * 4  # 16 px²
+        self.min_side = 3
+        self.max_aspect_ratio = 12.0
+
+        print(f"✅ ONNX model loaded from: {model_path}")
+        print(f"✅ ONNX providers: {self.session.get_providers()}")
+        print(f"✅ ONNX input: name={self.input_name}, shape={self.input_shape}")
 
     def __repr__(self) -> str:
-        return f"YOLOv26 Car Wash Miner classes={len(self.class_names)}"
+        return (
+            f"ONNXRuntime(session={type(self.session).__name__}, "
+            f"providers={self.session.get_providers()})"
+        )
 
     @staticmethod
     def _safe_dim(value, default: int) -> int:
         return value if isinstance(value, int) and value > 0 else default
 
-    # ─── Preprocessing ────────────────────────────────────────────
-    
+    @staticmethod
+    def _resolve_model_path(repo: Path) -> Path:
+        """Locate the ONNX model in the repo dir.
+
+        Prefers weights.onnx (FP16/FP32 export), then weights_int8.onnx (the
+        training script's INT8-quantized export -- works as-is: quantization
+        preserves the Ultralytics metadata and QDQ models take regular fp32
+        input), then any other .onnx file. INT8 is the fallback when the FP16
+        export exceeds the 30 MB deployment limit (e.g. yolo26m).
+        """
+        for name in ("weights.onnx", "weights_int8.onnx"):
+            p = repo / name
+            if p.exists():
+                if name != "weights.onnx":
+                    print(f"model: weights.onnx not found, using {name}")
+                return p
+        candidates = sorted(repo.glob("*.onnx"))
+        if candidates:
+            print(f"model: using {candidates[0].name}")
+            return candidates[0]
+        return repo / "weights.onnx"  # let session creation raise the error
+
+    def _read_model_class_order(self) -> list[str] | None:
+        """Read the model's class order from Ultralytics ONNX metadata.
+
+        Returns the class names ordered by model-emit index, or None when
+        metadata is missing/unparsable or doesn't match `class_names` as a
+        set (in which case the static _model_class_order fallback is used).
+        """
+        try:
+            import ast
+
+            meta = self.session.get_modelmeta().custom_metadata_map
+            names = ast.literal_eval(meta["names"])  # e.g. {0: 'broom', ...}
+            if isinstance(names, dict):
+                order = [str(names[i]) for i in sorted(names)]
+            else:
+                order = [str(n) for n in names]
+        except Exception as e:
+            print(f"cls order: could not read ONNX names metadata ({e})")
+            return None
+        if sorted(order) != sorted(self.class_names):
+            print(
+                f"cls order: ONNX names {order} do not match expected classes "
+                f"{self.class_names}; ignoring metadata"
+            )
+            return None
+        return order
+
     def _letterbox(
-        self, image: ndarray, new_shape: tuple[int, int],
-        color: tuple[int, int, int] = (114, 114, 114),
-    ) -> tuple[ndarray, float, float, float]:
+        self,
+        image: ndarray,
+        new_shape: tuple[int, int],
+        color=(114, 114, 114),
+    ) -> tuple[ndarray, float, tuple[float, float]]:
+        """
+        Resize with unchanged aspect ratio and pad to target shape.
+        Returns:
+            padded_image,
+            ratio,
+            (pad_w, pad_h)  # half-padding
+        """
+        h, w = image.shape[:2]
+        new_w, new_h = new_shape
+
+        ratio = min(new_w / w, new_h / h)
+        resized_w = int(round(w * ratio))
+        resized_h = int(round(h * ratio))
+
+        if (resized_w, resized_h) != (w, h):
+            interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR
+            image = cv2.resize(image, (resized_w, resized_h), interpolation=interp)
+
+        dw = new_w - resized_w
+        dh = new_h - resized_h
+        dw /= 2.0
+        dh /= 2.0
+
+        left = int(round(dw - 0.1))
+        right = int(round(dw + 0.1))
+        top = int(round(dh - 0.1))
+        bottom = int(round(dh + 0.1))
+
+        padded = cv2.copyMakeBorder(
+            image,
+            top,
+            bottom,
+            left,
+            right,
+            borderType=cv2.BORDER_CONSTANT,
+            value=color,
+        )
+        return padded, ratio, (dw, dh)
+
+    def _preprocess(
+        self, image: ndarray
+    ) -> tuple[np.ndarray, float, tuple[float, float], tuple[int, int]]:
+        """
+        Preprocess for fixed-size ONNX export:
+        - enhance image quality (CLAHE, denoise, sharpen)
+        - letterbox to model input size
+        - BGR -> RGB
+        - normalize to [0,1]
+        - HWC -> NCHW float32
+        """
         orig_h, orig_w = image.shape[:2]
-        target_w, target_h = new_shape
-        
-        r = min(target_w / orig_w, target_h / orig_h)
-        new_unpad_w = int(round(orig_w * r))
-        new_unpad_h = int(round(orig_h * r))
-        
-        resized = cv2.resize(image, (new_unpad_w, new_unpad_h), interpolation=cv2.INTER_LINEAR)
-        
-        dw = target_w - new_unpad_w
-        dh = target_h - new_unpad_h
-        pad_w = dw / 2.0
-        pad_h = dh / 2.0
-        
-        left = int(round(pad_w - 0.1))
-        right = int(round(pad_w + 0.1))
-        top = int(round(pad_h - 0.1))
-        bottom = int(round(pad_h + 0.1))
-        
-        out = cv2.copyMakeBorder(
-            resized, top, bottom, left, right,
-            cv2.BORDER_CONSTANT, value=color,
+
+        img, ratio, pad = self._letterbox(
+            image, (self.input_width, self.input_height)
         )
-        return out, r, pad_w, pad_h
-
-    def _preprocess(self, image_bgr: np.ndarray,
-                   allow_pad: bool = True) -> tuple[np.ndarray, dict]:
-        orig_h, orig_w = image_bgr.shape[:2]
-        extra_left = 0
-        extra_right = 0
-        if allow_pad and orig_w == orig_h:          # only pad when allowed
-            target_w = int(orig_w * 1.05)
-            if target_w > orig_w:
-                total_extra = target_w - orig_w
-                extra_left = total_extra // 2
-                extra_right = total_extra - extra_left
-                image_bgr = cv2.copyMakeBorder(
-                    image_bgr, 0, 0, extra_left, extra_right,
-                    cv2.BORDER_CONSTANT, value=(114, 114, 114),
-                )
-        padded_h, padded_w = image_bgr.shape[:2]
-        rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
-        img, ratio, pad_w, pad_h = self._letterbox(rgb, (self.input_w, self.input_h))
-        x = img.astype(np.float32) / 255.0
-        x = np.transpose(x, (2, 0, 1))[None, ...]
-        x = np.ascontiguousarray(x)
-        return x, {
-            "orig_h": orig_h, "orig_w": orig_w,
-            "ratio": ratio, "pad_w": pad_w, "pad_h": pad_h,
-            "extra_left": extra_left, "extra_right": extra_right,
-            "padded_w": padded_w, "padded_h": padded_h,
-        }
-
-    # ─── Vectorized box operations ───────────────────────────────
-    
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        img = (img.astype(np.float32) / 255.0)
+        img = np.transpose(img, (2, 0, 1))[None, ...]
+        img = np.ascontiguousarray(img, dtype=self.np_dtype)
+
+        return img, ratio, pad, (orig_w, orig_h)
+
     @staticmethod
     def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
         w, h = image_size
@@ -169,116 +282,200 @@ class Miner:
         return boxes
 
     @staticmethod
-    def _hard_nms(boxes: np.ndarray, scores: np.ndarray,
-                  iou_thresh: float) -> np.ndarray:
-        """Vectorized NMS. Returns indices to keep."""
-        n = len(boxes)
-        if n == 0:
-            return np.array([], dtype=np.intp)
-        order = np.argsort(-scores)
-        keep = []
-        while len(order) > 0:
-            i = int(order[0])
-            keep.append(i)
-            if len(order) == 1:
+    def _xywh_to_xyxy(boxes: np.ndarray) -> np.ndarray:
+        out = np.empty_like(boxes)
+        out[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.0
+        out[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.0
+        out[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.0
+        out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
+        return out
+
+    def _soft_nms(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        sigma: float = 0.5,
+        score_thresh: float = 0.01,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """
+        Soft-NMS: Gaussian decay of overlapping scores instead of hard removal.
+        Returns (kept_original_indices, updated_scores).
+        """
+        N = len(boxes)
+        if N == 0:
+            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
+
+        boxes = boxes.astype(np.float32, copy=True)
+        scores = scores.astype(np.float32, copy=True)
+        order = np.arange(N)
+
+        for i in range(N):
+            max_pos = i + int(np.argmax(scores[i:]))
+            boxes[[i, max_pos]] = boxes[[max_pos, i]]
+            scores[[i, max_pos]] = scores[[max_pos, i]]
+            order[[i, max_pos]] = order[[max_pos, i]]
+
+            if i + 1 >= N:
                 break
-            rest = order[1:]
-            xx1 = np.maximum(boxes[i, 0], boxes[rest, 0])
-            yy1 = np.maximum(boxes[i, 1], boxes[rest, 1])
-            xx2 = np.minimum(boxes[i, 2], boxes[rest, 2])
-            yy2 = np.minimum(boxes[i, 3], boxes[rest, 3])
+
+            xx1 = np.maximum(boxes[i, 0], boxes[i + 1:, 0])
+            yy1 = np.maximum(boxes[i, 1], boxes[i + 1:, 1])
+            xx2 = np.minimum(boxes[i, 2], boxes[i + 1:, 2])
+            yy2 = np.minimum(boxes[i, 3], boxes[i + 1:, 3])
             inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            a_i = (max(0.0, boxes[i, 2] - boxes[i, 0]) *
-                   max(0.0, boxes[i, 3] - boxes[i, 1]))
-            a_r = (np.maximum(0.0, boxes[rest, 2] - boxes[rest, 0]) *
-                   np.maximum(0.0, boxes[rest, 3] - boxes[rest, 1]))
-            iou = inter / (a_i + a_r - inter + 1e-7)
-            order = rest[iou <= iou_thresh]
-        return np.array(keep, dtype=np.intp)
-
-    def _per_class_hard_nms(self, boxes: np.ndarray, scores: np.ndarray,
-                            cls_ids: np.ndarray) -> np.ndarray:
-        """Per-class NMS using per-class IoU thresholds."""
+
+            area_i = max(0.0, float(
+                (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
+            ))
+            areas_j = (
+                np.maximum(0.0, boxes[i + 1:, 2] - boxes[i + 1:, 0])
+                * np.maximum(0.0, boxes[i + 1:, 3] - boxes[i + 1:, 1])
+            )
+            iou = inter / (area_i + areas_j - inter + 1e-7)
+            scores[i + 1:] *= np.exp(-(iou ** 2) / sigma)
+
+        mask = scores > score_thresh
+        return order[mask], scores[mask]
+
+    @staticmethod
+    def _hard_nms(
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        """
+        Standard NMS: keep one box per overlapping cluster (the one with highest score).
+        Returns indices of kept boxes (into the boxes/scores arrays).
+        """
+        N = len(boxes)
+        if N == 0:
+            return np.array([], dtype=np.intp)
+        boxes = np.asarray(boxes, dtype=np.float32)
+        scores = np.asarray(scores, dtype=np.float32)
+        order = np.argsort(scores)[::-1]
+        keep: list[int] = []
+        suppressed = np.zeros(N, dtype=bool)
+        for i in range(N):
+            idx = order[i]
+            if suppressed[idx]:
+                continue
+            keep.append(idx)
+            bi = boxes[idx]
+            for k in range(i + 1, N):
+                jdx = order[k]
+                if suppressed[jdx]:
+                    continue
+                bj = boxes[jdx]
+                xx1 = max(bi[0], bj[0])
+                yy1 = max(bi[1], bj[1])
+                xx2 = min(bi[2], bj[2])
+                yy2 = min(bi[3], bj[3])
+                inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1)
+                area_i = (bi[2] - bi[0]) * (bi[3] - bi[1])
+                area_j = (bj[2] - bj[0]) * (bj[3] - bj[1])
+                iou = inter / (area_i + area_j - inter + 1e-7)
+                if iou > iou_thresh:
+                    suppressed[jdx] = True
+        return np.array(keep)
+
+    def _per_class_hard_nms(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        """Hard NMS applied independently per class."""
         if len(boxes) == 0:
             return np.array([], dtype=np.intp)
-        all_keep = []
+        all_keep: list[int] = []
         for c in np.unique(cls_ids):
             mask = cls_ids == c
             indices = np.where(mask)[0]
-            cls_iou = float(self._iou_thres_array[c])  # per-class IoU threshold
-            keep = self._hard_nms(boxes[mask], scores[mask], cls_iou)
+            keep = self._hard_nms(boxes[mask], scores[mask], iou_thresh)
             all_keep.extend(indices[keep].tolist())
         all_keep.sort()
         return np.array(all_keep, dtype=np.intp)
 
-    def _cross_class_dedup_op(self, boxes: np.ndarray, scores: np.ndarray,
-                              cls_ids: np.ndarray, iou_thresh: float
-                              ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        n = len(boxes)
-        if n <= 1:
-            return boxes, scores, cls_ids
-        boxes = np.asarray(boxes, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
-        cls_ids = np.asarray(cls_ids, dtype=np.int32)
-        areas = (np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) *
-                 np.maximum(0.0, boxes[:, 3] - boxes[:, 1]))
-        margins = scores - self._conf_thres_array[cls_ids]
-        order = np.lexsort((-areas, -margins))
-        suppressed = np.zeros(n, dtype=bool)
-        keep = []
-        for i in order:
-            if suppressed[i]:
-                continue
-            keep.append(int(i))
-            bi = boxes[i]
-            xx1 = np.maximum(bi[0], boxes[:, 0])
-            yy1 = np.maximum(bi[1], boxes[:, 1])
-            xx2 = np.minimum(bi[2], boxes[:, 2])
-            yy2 = np.minimum(bi[3], boxes[:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            a_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
-            iou = inter / (a_i + areas - inter + 1e-7)
-            dup = iou > iou_thresh
-            dup[i] = False
-            suppressed |= dup
-        keep_idx = np.array(keep, dtype=np.intp)
-        return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx]
-
-    def _filter_sane_boxes(self, boxes: np.ndarray, scores: np.ndarray,
-                       cls_ids: np.ndarray, orig_size: tuple[int, int]
-                       ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """Filter by per-class min area, max area ratio, and aspect ratio."""
+    def _per_class_soft_nms(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        sigma: float = 0.5,
+        score_thresh: float = 0.01,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """Soft NMS applied independently per class."""
+        if len(boxes) == 0:
+            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
+        all_keep: list[int] = []
+        all_scores: list[float] = []
+        for c in np.unique(cls_ids):
+            mask = cls_ids == c
+            indices = np.where(mask)[0]
+            keep, updated = self._soft_nms(boxes[mask], scores[mask], sigma, score_thresh)
+            for k, s in zip(keep, updated):
+                all_keep.append(int(indices[k]))
+                all_scores.append(float(s))
+        if not all_keep:
+            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
+        return np.array(all_keep, dtype=np.intp), np.array(all_scores, dtype=np.float32)
+
+    def _filter_sane_boxes(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        orig_size: tuple[int, int],
+    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Filter out tiny, degenerate, or implausible boxes (common FP)."""
         if len(boxes) == 0:
             return boxes, scores, cls_ids
-        
         orig_w, orig_h = orig_size
         image_area = float(orig_w * orig_h)
-        bw = np.maximum(0.0, boxes[:, 2] - boxes[:, 0])
-        bh = np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
-        area = bw * bh
-        
-        ar = np.where(
-            (bw > 0) & (bh > 0),
-            np.maximum(bw / np.maximum(bh, 1e-6), bh / np.maximum(bw, 1e-6)),
-            np.inf,
-        )
-        
-        # Per-class minimum area
-        class_min_area = self._min_box_area_array[cls_ids]
-        
-        keep = (
-            (area >= class_min_area) &
-            (area <= 0.95 * image_area)
-        )
-        return boxes[keep], scores[keep], cls_ids[keep]
-
-    def _max_score_per_cluster(self, post_boxes: np.ndarray,
-                               post_cls: np.ndarray,
-                               full_boxes: np.ndarray,
-                               full_scores: np.ndarray,
-                               full_cls: np.ndarray,
-                               iou_thresh: float) -> np.ndarray:
-        """For each kept box, set confidence to max score in its SAME-CLASS cluster."""
+        keep = []
+        for i, box in enumerate(boxes):
+            x1, y1, x2, y2 = box.tolist()
+            bw = x2 - x1
+            bh = y2 - y1
+            if bw <= 0 or bh <= 0:
+                continue
+            if bw < self.min_side or bh < self.min_side:
+                continue
+            area = bw * bh
+            if area < self.min_box_area:
+                continue
+            if area > 0.95 * image_area:
+                continue
+            ar = max(bw / max(bh, 1e-6), bh / max(bw, 1e-6))
+            if ar > self.max_aspect_ratio:
+                continue
+            keep.append(i)
+        if not keep:
+            return (
+                np.empty((0, 4), dtype=np.float32),
+                np.empty((0,), dtype=np.float32),
+                np.empty((0,), dtype=np.int32),
+            )
+        k = np.array(keep, dtype=np.intp)
+        return boxes[k], scores[k], cls_ids[k]
+
+    @staticmethod
+    def _max_score_per_cluster(
+        post_boxes: np.ndarray,
+        post_cls: np.ndarray,
+        full_boxes: np.ndarray,
+        full_scores: np.ndarray,
+        full_cls: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        """For each kept (post-NMS) box, return the max score over the FULL
+        candidate set among SAME-CLASS boxes with IoU >= iou_thresh.
+
+        The previous version omitted the same-class constraint, which let a
+        confident broom raise the score of a coincident nozzle (or vice
+        versa) under TTA. That's a silent FP booster and is fixed here.
+        """
         n = len(post_boxes)
         if n == 0:
             return np.empty(0, dtype=np.float32)
@@ -298,16 +495,16 @@ class Miner:
             out[i] = float(np.max(full_scores[cluster])) if np.any(cluster) else 0.0
         return out
 
-    def _conf_filter_mask(self, scores: np.ndarray,
-                          cls_ids: np.ndarray, extra_left: int) -> np.ndarray:
-        """Per-class threshold with rescue bonus for missed classes."""
+    def _conf_filter_mask(
+        self, scores: np.ndarray, cls_ids: np.ndarray
+    ) -> np.ndarray:
+        """Boolean keep-mask: score >= per-class threshold, with a per-class
+        rescue -- if a class has zero boxes passing, admit its top-1 candidate
+        when its score >= (per-class threshold - per-class bonus).
+        """
         if len(scores) == 0:
             return np.zeros(0, dtype=bool)
-        thr = 0
-        if extra_left > 0:
-            thr = self._extra_conf_thres_array[cls_ids]
-        else:
-            thr = self._conf_thres_array[cls_ids]
+        thr = self._conf_thres_array[cls_ids]
         keep = scores >= thr
         for c in np.unique(cls_ids):
             b = float(self._bonus_array[c])
@@ -322,82 +519,60 @@ class Miner:
                 keep[top] = True
         return keep
 
-    def _suppress_overlapping_same_class(
+    def _cross_class_dedup_op(
         self,
         boxes: np.ndarray,
         scores: np.ndarray,
         cls_ids: np.ndarray,
-        threshold: float,
+        iou_thresh: float,
     ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """
-        Drop a same-class box that is (almost) entirely *contained* inside a larger
-        same-class box — a duplicate detection of one object.
-
-        Containment is intersection / area_of_SMALLER_box (IoMin), NOT IoU.
-        A small box nested in a large one has tiny IoU, so plain NMS never removes
-        it; IoMin catches it.
+        """Remove near-duplicate boxes across classes.
 
-            black (large) + green (fully inside black)  -> same gate, drop green
-            red   (sticks out of black)                 -> separate gate, keep
-
-        Survivor = the LARGER box, and its confidence is raised to the cluster max.
+        Order candidates by (score - per_class_threshold) margin, then by area;
+        keep the highest, suppress every other box with IoU > iou_thresh. For
+        car-wash this kills the common failure where water spray makes the
+        model fire both `nozzle` and `track` on the same patch, or where a
+        broom handle overlaps a drainage-gate detection.
         """
         n = len(boxes)
         if n <= 1:
             return boxes, scores, cls_ids
-
         boxes = np.asarray(boxes, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32).copy()
+        scores = np.asarray(scores, dtype=np.float32)
         cls_ids = np.asarray(cls_ids, dtype=np.int32)
-
         areas = (np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) *
-                np.maximum(0.0, boxes[:, 3] - boxes[:, 1]))
-
-        keep = np.ones(n, dtype=bool)
-
-        # Largest first, so the survivor of a containment chain is the biggest box.
-        order = np.argsort(-areas)
-
-        for idx_a in range(n):
-            a = order[idx_a]
-            if not keep[a]:
+                 np.maximum(0.0, boxes[:, 3] - boxes[:, 1]))
+        margins = scores - self._conf_thres_array[cls_ids]
+        order = np.lexsort((-areas, -margins))
+        suppressed = np.zeros(n, dtype=bool)
+        keep: list[int] = []
+        for i in order:
+            if suppressed[i]:
                 continue
-            for idx_b in range(idx_a + 1, n):
-                b = order[idx_b]              # areas[b] <= areas[a]
-                if not keep[b]:
-                    continue
-                if cls_ids[a] != cls_ids[b]:
-                    continue
-
-                x1 = max(boxes[a, 0], boxes[b, 0])
-                y1 = max(boxes[a, 1], boxes[b, 1])
-                x2 = min(boxes[a, 2], boxes[b, 2])
-                y2 = min(boxes[a, 3], boxes[b, 3])
-                if x2 <= x1 or y2 <= y1:
-                    continue
-                inter = (x2 - x1) * (y2 - y1)
-
-                # How much of the SMALLER box (b) lies inside the larger (a):
-                containment_b = inter / max(areas[b], 1e-9)
-
-                if containment_b >= threshold:        # b is nested -> it's a duplicate
-                    scores[a] = max(scores[a], scores[b])   # keep the higher score
-                    keep[b] = False                          # drop the smaller (green)
-
-        keep_idx = np.where(keep)[0]
+            keep.append(int(i))
+            bi = boxes[i]
+            xx1 = np.maximum(bi[0], boxes[:, 0])
+            yy1 = np.maximum(bi[1], boxes[:, 1])
+            xx2 = np.minimum(bi[2], boxes[:, 2])
+            yy2 = np.minimum(bi[3], boxes[:, 3])
+            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            a_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
+            iou = inter / (a_i + areas - inter + 1e-7)
+            dup = iou > iou_thresh
+            dup[i] = False
+            suppressed |= dup
+        keep_idx = np.array(keep, dtype=np.intp)
         return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx]
 
-    def _per_view_pipeline(self, boxes: np.ndarray, scores: np.ndarray,
-                           cls_ids: np.ndarray, orig_size: tuple[int, int]
-                           ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """Sanity filter + per-class NMS + cross-class dedup."""
-        boxes, scores, cls_ids = self._filter_sane_boxes(
-            boxes, scores, cls_ids, orig_size
-        )
-        if len(boxes) == 0:
-            return boxes, scores, cls_ids
+    def _per_view_pipeline(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Per-view post-processing: per-class NMS -> cap -> cross-class dedup."""
         if len(boxes) > 1:
-            keep = self._per_class_hard_nms(boxes, scores, cls_ids)
+            keep = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
             boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
         if len(scores) > self.max_det:
             top = np.argsort(-scores)[: self.max_det]
@@ -408,36 +583,32 @@ class Miner:
             )
         return boxes, scores, cls_ids
 
-    # ─── v26-specific decoding ────────────────────────────────────
-    
-    def _decode_v26_output(self, preds: np.ndarray, ratio: float,
-                           pad: tuple[float, float],
-                           orig_size: tuple[int, int],
-                           extra: tuple[int, int] = (0, 0)) -> list[BoundingBox]:  # NEW arg
-        """Decode YOLOv26 output (shape [1, 300, 6] or [300, 6])."""
+    def _decode_final_dets(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+        apply_optional_dedup: bool = False,
+    ) -> list[BoundingBox]:
+        """
+        Primary path:
+        expected output rows like [x1, y1, x2, y2, conf, cls_id]
+        in letterboxed input coordinates.
+        """
         if preds.ndim == 3 and preds.shape[0] == 1:
             preds = preds[0]
 
         if preds.ndim != 2 or preds.shape[1] < 6:
-            print(f"Warning: Unexpected v26 output shape: {preds.shape}")
-            return []
+            raise ValueError(f"Unexpected ONNX final-det output shape: {preds.shape}")
 
         boxes = preds[:, :4].astype(np.float32)
         scores = preds[:, 4].astype(np.float32)
         cls_ids = preds[:, 5].astype(np.int32)
+        cls_ids = self.cls_remap[cls_ids]
 
-        n_cls = len(self.class_names)
-        valid = (cls_ids >= 0) & (cls_ids < n_cls)
-        boxes = boxes[valid]
-        scores = scores[valid]
-        cls_ids = cls_ids[valid]
-
-        if len(boxes) == 0:
-            return []
-
-        extra_left, _extra_right = extra
-
-        keep = self._conf_filter_mask(scores, cls_ids, extra_left)
+        # Per-class confidence filter with rescue (replaces scalar threshold)
+        keep = self._conf_filter_mask(scores, cls_ids)
         boxes = boxes[keep]
         scores = scores[keep]
         cls_ids = cls_ids[keep]
@@ -445,169 +616,262 @@ class Miner:
         if len(boxes) == 0:
             return []
 
-        # 1) undo letterbox -> coords in the PADDED (widened) image
         pad_w, pad_h = pad
+        orig_w, orig_h = orig_size
+
+        # reverse letterbox
         boxes[:, [0, 2]] -= pad_w
         boxes[:, [1, 3]] -= pad_h
         boxes /= ratio
+        boxes = self._clip_boxes(boxes, (orig_w, orig_h))
 
-        # 2) NEW: undo the left/right pre-padding -> coords in the ORIGINAL image.
-        #    Only the LEFT pad shifts x; right pad adds width but no offset.
-        
-        if extra_left:
-            boxes[:, [0, 2]] -= extra_left
-        
-        # 2b) NEW: drop boxes that fall in the black padding region.
-        #     A real detection must have its CENTER inside the original image
-        #     width [0, orig_w]; boxes centered in the black bars are spurious.
-        if extra_left or _extra_right:
-            orig_w, orig_h = orig_size
-            cx = (boxes[:, 0] + boxes[:, 2]) * 0.5
-            inside = (cx >= 0) & (cx <= orig_w)
-            boxes = boxes[inside]
-            scores = scores[inside]
-            cls_ids = cls_ids[inside]
-            if len(boxes) == 0:
-                return []
-
-        # 3) clip to ORIGINAL image bounds (orig_size is the true original size)
-        boxes = self._clip_boxes(boxes, orig_size)
-
-        boxes, scores, cls_ids = self._per_view_pipeline(
+        # Box sanity filter (reduces FP)
+        boxes, scores, cls_ids = self._filter_sane_boxes(
             boxes, scores, cls_ids, orig_size
         )
+        if len(boxes) == 0:
+            return []
 
-        return self._build_results(boxes, scores, cls_ids, orig_size)
+        if apply_optional_dedup and len(boxes) > 1:
+            # Soft-NMS path preserved as a tunable option; default below.
+            keep_idx, scores = self._per_class_soft_nms(boxes, scores, cls_ids)
+            boxes = boxes[keep_idx]
+            cls_ids = cls_ids[keep_idx]
+            if len(scores) > self.max_det:
+                top = np.argsort(-scores)[: self.max_det]
+                boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]
+            if len(boxes) > 1:
+                boxes, scores, cls_ids = self._cross_class_dedup_op(
+                    boxes, scores, cls_ids, self.cross_iou_thresh
+                )
+        else:
+            # Default: per-class hard NMS -> cap -> cross-class dedup
+            boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
 
-    @staticmethod
-    def _build_results(boxes: np.ndarray, scores: np.ndarray,
-                       cls_ids: np.ndarray,
-                       orig_size: tuple[int, int]) -> list[BoundingBox]:
-        results = []
+        results: list[BoundingBox] = []
+        for box, conf, cls_id in zip(boxes, scores, cls_ids):
+            x1, y1, x2, y2 = box.tolist()
+
+            if x2 <= x1 or y2 <= y1:
+                continue
+
+            results.append(
+                BoundingBox(
+                    x1=int(math.floor(x1)),
+                    y1=int(math.floor(y1)),
+                    x2=int(math.ceil(x2)),
+                    y2=int(math.ceil(y2)),
+                    cls_id=int(cls_id),
+                    conf=float(conf),
+                )
+            )
+
+        return results
+
+    def _decode_raw_yolo(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """
+        Fallback path for raw YOLO predictions.
+        Supports common layouts:
+        - [1, C, N]
+        - [1, N, C]
+        """
+        if preds.ndim != 3:
+            raise ValueError(f"Unexpected raw ONNX output shape: {preds.shape}")
+
+        if preds.shape[0] != 1:
+            raise ValueError(f"Unexpected batch dimension in raw output: {preds.shape}")
+
+        preds = preds[0]
+
+        # Normalize to [N, C]
+        if preds.shape[0] <= 16 and preds.shape[1] > preds.shape[0]:
+            preds = preds.T
+
+        if preds.ndim != 2 or preds.shape[1] < 5:
+            raise ValueError(f"Unexpected normalized raw output shape: {preds.shape}")
+
+        boxes_xywh = preds[:, :4].astype(np.float32)
+        cls_part = preds[:, 4:].astype(np.float32)
+
+        if cls_part.shape[1] == 1:
+            scores = cls_part[:, 0]
+            cls_ids = np.zeros(len(scores), dtype=np.int32)
+        else:
+            cls_ids = np.argmax(cls_part, axis=1).astype(np.int32)
+            scores = cls_part[np.arange(len(cls_part)), cls_ids]
+        cls_ids = self.cls_remap[cls_ids]
+
+        # Per-class confidence filter with rescue (replaces scalar threshold)
+        keep = self._conf_filter_mask(scores, cls_ids)
+        boxes_xywh = boxes_xywh[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+        if len(boxes_xywh) == 0:
+            return []
+
+        boxes = self._xywh_to_xyxy(boxes_xywh)
+
+        # Order matches fire001 / _decode_final_dets:
+        # unscale -> clip -> sanity filter -> per-view pipeline (NMS, cap, cross-class dedup).
+        pad_w, pad_h = pad
         orig_w, orig_h = orig_size
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, (orig_w, orig_h))
+
+        boxes, scores, cls_ids = self._filter_sane_boxes(
+            boxes, scores, cls_ids, (orig_w, orig_h)
+        )
+        if len(boxes) == 0:
+            return []
+
+        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
+
+        results: list[BoundingBox] = []
         for box, conf, cls_id in zip(boxes, scores, cls_ids):
             x1, y1, x2, y2 = box.tolist()
+
             if x2 <= x1 or y2 <= y1:
                 continue
+
             results.append(
                 BoundingBox(
-                    x1=max(0, min(orig_w, int(math.floor(x1)))),
-                    y1=max(0, min(orig_h, int(math.floor(y1)))),
-                    x2=max(0, min(orig_w, int(math.ceil(x2)))),
-                    y2=max(0, min(orig_h, int(math.ceil(y2)))),
+                    x1=int(math.floor(x1)),
+                    y1=int(math.floor(y1)),
+                    x2=int(math.ceil(x2)),
+                    y2=int(math.ceil(y2)),
                     cls_id=int(cls_id),
-                    conf=float(max(0.0, min(1.0, conf))),
+                    conf=float(conf),
                 )
             )
+
         return results
 
-    # ─── Single-view inference ────────────────────────────────────
-    
-    def _predict_single(self, image_bgr: np.ndarray,
-                        allow_pad: bool = True) -> list[BoundingBox]:
-        if image_bgr is None or not isinstance(image_bgr, np.ndarray):
-            raise ValueError("Invalid image input")
-        if image_bgr.dtype != np.uint8:
-            image_bgr = image_bgr.astype(np.uint8)
-
-        inp, meta = self._preprocess(image_bgr, allow_pad=allow_pad)
-        outputs = self.session.run(None, {self.input_name: inp})
-
-        ratio = float(meta["ratio"])
-        pad = (float(meta["pad_w"]), float(meta["pad_h"]))
-        orig_size = (int(meta["orig_w"]), int(meta["orig_h"]))
-        extra = (int(meta["extra_left"]), int(meta["extra_right"]))
-
-        return self._decode_v26_output(outputs[0], ratio, pad, orig_size, extra)
-
-    # ─── TTA inference ────────────────────────────────────────────
-    
-    def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
-        """3-view TTA: original (no pad) + flip (no pad) + original (L/R padded).
-        All three views return ORIGINAL-image coords, then pooled."""
-        orig_h, orig_w = image_bgr.shape[:2]
-
-        # View 1: original, NO left/right padding
-        boxes_orig = self._predict_single(image_bgr, allow_pad=True)
-        
-        # View 2: horizontal flip, NO left/right padding
-        flipped = cv2.flip(image_bgr, 1)
-        boxes_flip = self._predict_single(flipped, allow_pad=True)
-        w = image_bgr.shape[1]
+    def _postprocess(
+        self,
+        output: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """
+        Prefer final detections first.
+        Fallback to raw decode only if needed.
+        """
+        # final detections: [N,6]
+        if output.ndim == 2 and output.shape[1] >= 6:
+            return self._decode_final_dets(output, ratio, pad, orig_size)
+
+        # final detections: [1,N,6]
+        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 6:
+            return self._decode_final_dets(output, ratio, pad, orig_size)
+
+        # fallback raw decode
+        return self._decode_raw_yolo(output, ratio, pad, orig_size)
+
+    def _predict_single(self, image: np.ndarray) -> list[BoundingBox]:
+        if image is None:
+            raise ValueError("Input image is None")
+        if not isinstance(image, np.ndarray):
+            raise TypeError(f"Input is not numpy array: {type(image)}")
+        if image.ndim != 3:
+            raise ValueError(f"Expected HWC image, got shape={image.shape}")
+        if image.shape[0] <= 0 or image.shape[1] <= 0:
+            raise ValueError(f"Invalid image shape={image.shape}")
+        if image.shape[2] != 3:
+            raise ValueError(f"Expected 3 channels, got shape={image.shape}")
+
+        if image.dtype != np.uint8:
+            image = image.astype(np.uint8)
+
+        input_tensor, ratio, pad, orig_size = self._preprocess(image)
+
+        expected_shape = (1, 3, self.input_height, self.input_width)
+        if input_tensor.shape != expected_shape:
+            raise ValueError(
+                f"Bad input tensor shape={input_tensor.shape}, expected={expected_shape}"
+            )
+
+        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
+        det_output = outputs[0]
+        return self._postprocess(det_output, ratio, pad, orig_size)
+
+    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        """Horizontal-flip TTA.
+
+        Strategy (ported from fire001):
+          1. Predict on original and on flipped image.
+          2. Map flipped boxes back to original coordinates.
+          3. Per-class hard NMS on the union.
+          4. For each kept box, compute the max SAME-CLASS score across the
+             FULL union -- a high-confidence flipped detection raises a
+             borderline original one, but never one of a different class.
+          5. Cross-class dedup to suppress same-physical-object multi-class.
+        """
+        boxes_orig = self._predict_single(image)
+
+        flipped = cv2.flip(image, 1)
+        boxes_flip = self._predict_single(flipped)
+
+        w = image.shape[1]
         boxes_flip = [
-            BoundingBox(x1=w - b.x2, y1=b.y1, x2=w - b.x1, y2=b.y2,
-                        cls_id=b.cls_id, conf=b.conf)
+            BoundingBox(
+                x1=w - b.x2, y1=b.y1, x2=w - b.x1, y2=b.y2,
+                cls_id=b.cls_id, conf=b.conf,
+            )
             for b in boxes_flip
         ]
-        
-        # View 3: original, WITH left/right padding (only meaningful if square)
-        # boxes_pad = self._predict_single(image_bgr, allow_pad=True)
-        # NOZZLE_CLS = self.class_names.index('nozzle')   # == 2
-        # boxes_pad = [b for b in boxes_pad if b.cls_id != NOZZLE_CLS]
-
-        # # View 4: flip, WITH left/right padding (only meaningful if square)
-        # flipped = cv2.flip(image_bgr, 1)
-        # boxes_pad_flip = self._predict_single(flipped, allow_pad=True)
-        # boxes_pad_flip = [
-        #     BoundingBox(x1=w - b.x2, y1=b.y1, x2=w - b.x1, y2=b.y2,
-        #                 cls_id=b.cls_id, conf=b.conf)
-        #     for b in boxes_pad_flip
-        # ]
-        # NOZZLE_CLS = self.class_names.index('nozzle')   # == 2
-        # boxes_pad_flip = [b for b in boxes_pad_flip if b.cls_id != NOZZLE_CLS]
-
-        all_boxes = boxes_orig + boxes_flip# + boxes_pad# + boxes_pad_flip
-        if not all_boxes:
+
+        all_boxes = boxes_orig + boxes_flip
+        if len(all_boxes) == 0:
             return []
 
-        coords = np.array([[b.x1, b.y1, b.x2, b.y2] for b in all_boxes], dtype=np.float32)
+        coords = np.array(
+            [[b.x1, b.y1, b.x2, b.y2] for b in all_boxes], dtype=np.float32
+        )
         scores = np.array([b.conf for b in all_boxes], dtype=np.float32)
         cls_ids = np.array([b.cls_id for b in all_boxes], dtype=np.int32)
 
-        # Per-class NMS (uses per-class IoU thresholds)
-        hard_keep = self._per_class_hard_nms(coords, scores, cls_ids)
+        hard_keep = self._per_class_hard_nms(coords, scores, cls_ids, self.iou_thres)
         if len(hard_keep) == 0:
             return []
-        
         if len(hard_keep) > self.max_det:
             top = np.argsort(-scores[hard_keep])[: self.max_det]
             hard_keep = hard_keep[top]
-        
-        # For confidence boost, use average IoU threshold (or could use median)
-        avg_iou = float(np.mean(self._iou_thres_array))
+
+        # Class-aware cluster-max score boost (fixes the silent cross-class
+        # leak in the previous _max_score_per_cluster).
         boosted = self._max_score_per_cluster(
             coords[hard_keep], cls_ids[hard_keep],
-            coords, scores, cls_ids, avg_iou,
+            coords, scores, cls_ids, self.iou_thres,
         )
 
         kept_coords = coords[hard_keep]
         kept_cls = cls_ids[hard_keep]
-        
-        # Cross-class dedup
         if len(kept_coords) > 1:
             kept_coords, boosted, kept_cls = self._cross_class_dedup_op(
                 kept_coords, boosted, kept_cls, self.cross_iou_thresh
             )
 
-        out_boxes = []
-        for j in range(len(kept_coords)):
-            x1, y1, x2, y2 = kept_coords[j].tolist()
-            if x2 <= x1 or y2 <= y1:
-                continue
-            out_boxes.append(
-                BoundingBox(
-                    x1=max(0, min(orig_w, int(math.floor(x1)))),
-                    y1=max(0, min(orig_h, int(math.floor(y1)))),
-                    x2=max(0, min(orig_w, int(math.ceil(x2)))),
-                    y2=max(0, min(orig_h, int(math.ceil(y2)))),
-                    cls_id=int(kept_cls[j]),
-                    conf=float(max(0.0, min(1.0, boosted[j]))),
-                )
+        return [
+            BoundingBox(
+                x1=int(math.floor(kept_coords[j, 0])),
+                y1=int(math.floor(kept_coords[j, 1])),
+                x2=int(math.ceil(kept_coords[j, 2])),
+                y2=int(math.ceil(kept_coords[j, 3])),
+                cls_id=int(kept_cls[j]),
+                conf=float(boosted[j]),
             )
-        return out_boxes
+            for j in range(len(kept_coords))
+        ]
 
-    # ─── Public API ───────────────────────────────────────────────
-    
     def predict_batch(
         self,
         batch_images: list[ndarray],
@@ -615,18 +879,23 @@ class Miner:
         n_keypoints: int,
     ) -> list[TVFrameResult]:
         results: list[TVFrameResult] = []
-        for idx, image in enumerate(batch_images):
+
+        for frame_number_in_batch, image in enumerate(batch_images):
             try:
-                boxes = self._infer_single(image)
+                if self.use_tta:
+                    boxes = self._predict_tta(image)
+                else:
+                    boxes = self._predict_single(image)
             except Exception as e:
-                print(f"Inference failed for frame {offset + idx}: {e}")
+                print(f"⚠️ Inference failed for frame {offset + frame_number_in_batch}: {e}")
                 boxes = []
-            keypoints = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
+
             results.append(
                 TVFrameResult(
-                    frame_id=offset + idx,
+                    frame_id=offset + frame_number_in_batch,
                     boxes=boxes,
-                    keypoints=keypoints,
+                    keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
                 )
             )
+
         return results

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5eff40e23f79ec4d26d8639de704fdc432abab52f0fc44ec908037e9a2316824
-size 20833918
+oid sha256:7d0089998a8868844db5012f76470768e2cc92eeee92a2c22e84c46fd4109ea5
+size 19287011