Upload folder using huggingface_hub

miner.py

@@ -8,6 +8,35 @@ from numpy import ndarray
 from pydantic import BaseModel
 
 
+# =============================================================================
+# BEST merged car-wash miner.
+#
+# Base: carwash001/washvision01 (robust model loading + correct coordinate math)
+# Merged-in ideas from ScoreVisionCarWash (the one genuinely different rival):
+#   * vectorized _hard_nms                 -> O(n^2) numpy, not a Python loop
+#   * vectorized _max_score_per_cluster    -> single IoU matrix, not per-box loop
+#   * per-class NMS-IoU array              -> _iou_thres_array (was one global)
+#   * per-class min-area array             -> _min_box_area_array (was one global)
+#   * pre_nms_topk                         -> bounds NMS cost on crowded frames
+#   * single-view cluster boost            -> confidence recovery on the DEPLOYED
+#                                             (non-TTA) path, not just under TTA
+#   * square-domain hook                   -> the validator scores 1024x1024
+#                                             squished frames (orig_w == orig_h);
+#                                             detect that and (a) use a val-tuned
+#                                             threshold set, (b) optionally widen
+#                                             to partly de-squish.
+# Kept from the robust base (ScoreVision lacks all of these):
+#   * FP16 input auto-detect, cls_remap from ONNX metadata, INT8 + raw-YOLO
+#     decode fallbacks, ORT thread pinning, "no CLAHE" (train/test match).
+#
+# DEFAULTS reproduce carwash001's current DEPLOYED behavior (per-class IoU all
+# 0.5, cluster boost off, square-pad off, thresholds = the tuned set) so this is
+# a safe drop-in. The NEW levers are exposed but neutral until you re-tune them
+# on the true-domain val (car-wash-55-styled/valid) with tune_miner.py. Do NOT
+# assume the new knobs help before that sweep -- they are opportunities, measured.
+# =============================================================================
+
+
 class BoundingBox(BaseModel):
     x1: int
     y1: int
@@ -24,32 +53,26 @@ class TVFrameResult(BaseModel):
 
 
 class Miner:
-    def __init__(self,
-        path_hf_repo: Path
-    ) -> None:
+    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.
+        # Canonical class order every downstream consumer (validator,
+        # BoundingBox.cls_id) sees: 0=broom, 1=drainage gate, 2=nozzle, 3=track.
         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.
+        # Fallback model-emit order (used only when ONNX metadata is missing).
         self._model_class_order = ["broom", "drainage gate", "nozzle", "track"]
         print("ORT version:", ort.__version__)
 
         try:
             ort.preload_dlls()
-            print("✅ onnxruntime.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
+        # Pin threads for the CPU latency gate (ScoreVision leaves these default).
         sess_options.intra_op_num_threads = 2
         sess_options.inter_op_num_threads = 1
         sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
@@ -60,21 +83,18 @@ class Miner:
                 sess_options=sess_options,
                 providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
             )
-            print("✅ Created ORT session with preferred CUDA provider list")
+            print("Created ORT session with preferred CUDA provider list")
         except Exception as e:
-            print(f"⚠️ CUDA session creation failed, falling back to CPU: {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.
+        # cls_remap[i] = self.class_names.index(model_class_order[i]); order comes
+        # from ONNX metadata when present, else the static fallback.
         model_class_order = self._read_model_class_order()
         if model_class_order is None:
             model_class_order = list(self._model_class_order)
@@ -87,81 +107,84 @@ class Miner:
 
         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()]
+        self.output_names = [o.name for o in self.session.get_outputs()]
         self.input_shape = self.session.get_inputs()[0].shape
 
-        # Match the ONNX input dtype (this export is FP16 -> needs float16 input).
+        # Match the ONNX input dtype (FP16 export 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 TTA-off grid sweep
-        # on the validator-style val split (tune_miner.py, car-wash-49-val1024
-        # val, ALL 2476 crops, against the synthetic-crop model; composite
-        # 0.8716 -> 0.8731) -- 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)
+        print(f"ONNX input dtype: {input_type} -> numpy {self.np_dtype}")
+
+        # The miner MUST run at the ONNX's baked input size. Dropping this to
+        # 640 (ScoreVision) is a real latency lever but is an EXPORT decision
+        # (re-export + re-eval; risks nozzle recall), not a miner-side change.
+        self.input_height = self._safe_dim(self.input_shape[2], default=704)
+        self.input_width = self._safe_dim(self.input_shape[3], default=704)
+
+        # ── Post-processing config (tune on car-wash-55-styled/valid) ─────────
+        # Per-class NMS IoU. Default all 0.5 == carwash001's single global value.
+        # ScoreVision uses [0.6,0.7,0.5,0.7]; sweep before adopting.
+        self._iou_thres_array = np.array([0.5, 0.5, 0.5, 0.5], dtype=np.float32)
+        self.cross_iou_thresh = 0.9   # cross-class dedup IoU (same physical object, 2 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 = False
-
-#       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].
-        # Values are the 2476-crop TTA-off sweep optimum on the synthetic-crop
-        # model: lowering drainage (0.30->0.22) and nozzle (0.45->0.38) recovers
-        # recall (map50 0.836->0.847) while FP actually drops (0.925->0.913).
-        self._conf_thres_array = np.array(
-            [0.30, 0.30, 0.40, 0.30], 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 2476-crop sweep (rescue_bonus=False won)
-        # confirmed rescue admits more false positives than true positives
-        # under the validator's FP pillar.
-        self._bonus_array = np.array(
-            [0.05, 0.05, 0.1, 0.05], 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.pre_nms_topk = 1000      # cap candidates before NMS (crowded-frame speed guard)
+        self.use_tta = False          # 2nd flipped pass; doubles latency -> off for CPU gate
+
+        # Single-view cluster boost: raise each survivor's conf to its same-class
+        # IoU-cluster max. OFF by default (can raise FP under the FP pillar);
+        # ScoreVision runs it on. Sweep on the true-domain val before enabling.
+        self.use_cluster_boost = True
+
+        # Per-class confidence thresholds. `_conf_thres_array` is the SQUARE /
+        # validator-eval set (every scored frame is 1024x1024, so this is the
+        # one that matters); `_extra` is a fallback for non-square inputs
+        # (warmup / any 16:9 frame). Default: both = the tuned set.
+        # Current tuned optimum (2476-crop TTA-off sweep): 0.836->0.847 map50,
+        # FP 0.925->0.913. Re-run tune_miner on car-wash-55-styled/valid.
+        self._conf_thres_array = np.array([0.22, 0.22, 0.38, 0.28], dtype=np.float32)
+        self._extra_conf_thres_array = np.array([0.25, 0.25, 0.45, 0.30], dtype=np.float32)
+
+        # Per-class rescue bonus: if a class has ZERO boxes passing, admit its
+        # top-1 when score >= (threshold - bonus). DISABLED (all zeros): the
+        # sweep showed rescue admits more FP than TP under the FP pillar.
+        self._bonus_array = np.array([0.02, 0.02, 0.03, 0.03], dtype=np.float32)
+
+        # Per-class min box area (px^2). nozzle boxes are tiny (GT median ~290,
+        # min ~32 px^2) so its floor stays low. Default loose; ScoreVision uses
+        # [144,144,4,64]. Max-area cap is a fraction of the frame.
+        self._min_box_area_array = np.array([16.0, 16.0, 4.0, 16.0], dtype=np.float32)
         self.min_side = 3
         self.max_aspect_ratio = 12.0
+        self.max_area_frac = 0.95
 
-        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}")
+        # Square-domain de-squish: widen a square (validator) frame by this
+        # fraction before letterboxing, then drop pad-center boxes and un-pad.
+        # 0.0 = off (our model already scores 0.916 on squished val without it;
+        # enabling changes aspect -> validate first). ScoreVision uses 0.05.
+        self.square_pad_frac = 0.0
 
+        self._avg_iou = float(np.mean(self._iou_thres_array))
+        print(f"ONNX model loaded from: {model_path}")
+        print(f"ONNX input: name={self.input_name}, shape={self.input_shape}")
         self._warmup()
 
+    # ── Setup helpers ────────────────────────────────────────────────────────
+
     def _warmup(self, iters: int = 3) -> None:
         try:
             dummy = np.zeros((720, 1280, 3), dtype=np.uint8)
             for _ in range(max(1, iters)):
                 self.predict_batch(batch_images=[dummy], offset=0, n_keypoints=0)
-            print(f"✅ warmup: {iters} dummy predict_batch call(s) done")
+            print(f"warmup: {iters} dummy predict_batch call(s) done")
         except Exception as e:
-            print(f"⚠️ warmup skipped: {e}")
+            print(f"warmup skipped: {e}")
 
     def __repr__(self) -> str:
-        return (
-            f"ONNXRuntime(session={type(self.session).__name__}, "
-            f"providers={self.session.get_providers()})"
-        )
+        return f"CarWashMiner(classes={len(self.class_names)}, providers={self.session.get_providers()})"
 
     @staticmethod
     def _safe_dim(value, default: int) -> int:
@@ -169,14 +192,7 @@ class Miner:
 
     @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).
-        """
+        """Prefer weights.onnx, then weights_int8.onnx, then any .onnx."""
         for name in ("weights.onnx", "weights_int8.onnx"):
             p = repo / name
             if p.exists():
@@ -190,99 +206,72 @@ class Miner:
         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).
-        """
+        """Read class order from Ultralytics ONNX `names` metadata, or None."""
         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]
+            names = ast.literal_eval(meta["names"])
+            order = ([str(names[i]) for i in sorted(names)] if isinstance(names, dict)
+                     else [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"
-            )
+            print(f"cls order: ONNX names {order} != expected {self.class_names}; ignoring")
             return None
         return order
 
-    def _letterbox(
-        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
-        """
+    # ── Preprocessing ────────────────────────────────────────────────────────
+
+    def _letterbox(self, image: ndarray, new_shape: tuple[int, int],
+                   color=(114, 114, 114)) -> tuple[ndarray, float, tuple[float, float]]:
         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):
+        rw, rh = int(round(w * ratio)), int(round(h * ratio))
+        if (rw, rh) != (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,
-        )
+            image = cv2.resize(image, (rw, rh), interpolation=interp)
+        dw = (new_w - rw) / 2.0
+        dh = (new_h - rh) / 2.0
+        left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+        top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+        padded = cv2.copyMakeBorder(image, top, bottom, left, right,
+                                    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
+    def _preprocess(self, image: ndarray) -> tuple[np.ndarray, dict]:
+        """Letterbox to the ONNX input size; NO CLAHE/denoise/sharpen (any
+        enhancement the model wasn't trained on is a train/test mismatch that
+        HURTS accuracy and also risks the CPU latency gate).
+
+        Square-domain de-squish: when the frame is square (the validator's
+        1024x1024 squished eval image) and square_pad_frac > 0, widen it with
+        gray bars first -- objects come out horizontally compressed in those
+        frames, and mild widening partly restores their aspect.
         """
         orig_h, orig_w = image.shape[:2]
-
-        img, ratio, pad = self._letterbox(
-            image, (self.input_width, self.input_height)
-        )
+        extra_left = extra_right = 0
+        if self.square_pad_frac > 0.0 and orig_w == orig_h:
+            target_w = int(orig_w * (1.0 + self.square_pad_frac))
+            if target_w > orig_w:
+                total = target_w - orig_w
+                extra_left = total // 2
+                extra_right = total - extra_left
+                image = cv2.copyMakeBorder(image, 0, 0, extra_left, extra_right,
+                                           cv2.BORDER_CONSTANT, value=(114, 114, 114))
+
+        img, ratio, pad = self._letterbox(image, (self.input_width, self.input_height))
         img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-        img = (img.astype(np.float32) / 255.0)
+        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": ratio, "pad": pad, "orig_size": (orig_w, orig_h),
+            "extra_left": extra_left, "extra_right": extra_right,
+        }
 
-        return img, ratio, pad, (orig_w, orig_h)
+    # ── Vectorized box ops ───────────────────────────────────────────────────
 
     @staticmethod
     def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
@@ -302,222 +291,81 @@ class Miner:
         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:
+    @staticmethod
+    def _hard_nms(boxes: np.ndarray, scores: np.ndarray, iou_thresh: float) -> np.ndarray:
+        """Vectorized greedy NMS (ScoreVision): areas precomputed once, each
+        step is a single numpy IoU vector -- no inner Python loop."""
+        n = len(boxes)
+        if n == 0:
+            return np.array([], dtype=np.intp)
+        x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
+        areas = np.maximum(0.0, x2 - x1) * np.maximum(0.0, y2 - y1)
+        order = np.argsort(-scores)
+        keep = []
+        while order.size > 0:
+            i = int(order[0])
+            keep.append(i)
+            if order.size == 1:
                 break
-
-            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])
+            rest = order[1:]
+            xx1 = np.maximum(x1[i], x1[rest])
+            yy1 = np.maximum(y1[i], y1[rest])
+            xx2 = np.minimum(x2[i], x2[rest])
+            yy2 = np.minimum(y2[i], y2[rest])
             inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            iou = inter / (areas[i] + areas[rest] - inter + 1e-7)
+            order = rest[iou <= iou_thresh]
+        return np.array(keep, dtype=np.intp)
 
-            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."""
+    def _per_class_hard_nms(self, boxes: np.ndarray, scores: np.ndarray,
+                            cls_ids: np.ndarray) -> np.ndarray:
+        """Per-class NMS using the per-class IoU thresholds."""
         if len(boxes) == 0:
             return np.array([], dtype=np.intp)
         all_keep: list[int] = []
         for c in np.unique(cls_ids):
-            mask = cls_ids == c
-            indices = np.where(mask)[0]
-            keep = self._hard_nms(boxes[mask], scores[mask], iou_thresh)
-            all_keep.extend(indices[keep].tolist())
+            idx = np.where(cls_ids == c)[0]
+            cls_iou = float(self._iou_thres_array[c])
+            keep = self._hard_nms(boxes[idx], scores[idx], cls_iou)
+            all_keep.extend(idx[keep].tolist())
         all_keep.sort()
         return np.array(all_keep, dtype=np.intp)
 
-    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)
-        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.
-        """
+    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:
+        """Each survivor's confidence -> max score in its SAME-CLASS IoU cluster.
+        Vectorized single (n_post x n_full) IoU matrix (ScoreVision)."""
         n = len(post_boxes)
         if n == 0:
             return np.empty(0, dtype=np.float32)
-        full_areas = (np.maximum(0.0, full_boxes[:, 2] - full_boxes[:, 0]) *
-                      np.maximum(0.0, full_boxes[:, 3] - full_boxes[:, 1]))
-        out = np.empty(n, dtype=np.float32)
-        for i in range(n):
-            bi = post_boxes[i]
-            xx1 = np.maximum(bi[0], full_boxes[:, 0])
-            yy1 = np.maximum(bi[1], full_boxes[:, 1])
-            xx2 = np.minimum(bi[2], full_boxes[:, 2])
-            yy2 = np.minimum(bi[3], full_boxes[:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            a_i = max(0.0, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
-            iou = inter / (a_i + full_areas - inter + 1e-7)
-            cluster = (iou >= iou_thresh) & (full_cls == post_cls[i])
-            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
-    ) -> 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).
-        """
+        m = len(full_boxes)
+        if m == 0:
+            return np.zeros(n, dtype=np.float32)
+        pa = (np.maximum(0.0, post_boxes[:, 2] - post_boxes[:, 0]) *
+              np.maximum(0.0, post_boxes[:, 3] - post_boxes[:, 1]))
+        fa = (np.maximum(0.0, full_boxes[:, 2] - full_boxes[:, 0]) *
+              np.maximum(0.0, full_boxes[:, 3] - full_boxes[:, 1]))
+        xx1 = np.maximum(post_boxes[:, 0][:, None], full_boxes[:, 0][None, :])
+        yy1 = np.maximum(post_boxes[:, 1][:, None], full_boxes[:, 1][None, :])
+        xx2 = np.minimum(post_boxes[:, 2][:, None], full_boxes[:, 2][None, :])
+        yy2 = np.minimum(post_boxes[:, 3][:, None], full_boxes[:, 3][None, :])
+        inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+        iou = inter / (pa[:, None] + fa[None, :] - inter + 1e-7)
+        mask = (iou >= iou_thresh) & (post_cls[:, None] == full_cls[None, :])
+        tiled = np.where(mask, full_scores[None, :], -np.inf)
+        out = tiled.max(axis=1)
+        out[~np.isfinite(out)] = 0.0
+        return out.astype(np.float32)
+
+    def _conf_filter_mask(self, scores: np.ndarray, cls_ids: np.ndarray,
+                          extra_left: int) -> np.ndarray:
+        """Per-class threshold (square vs non-square set) + per-class rescue."""
         if len(scores) == 0:
             return np.zeros(0, dtype=bool)
-        thr = self._conf_thres_array[cls_ids]
-        keep = scores >= thr
+        thr_arr = self._extra_conf_thres_array if extra_left > 0 else self._conf_thres_array
+        keep = scores >= thr_arr[cls_ids]
         for c in np.unique(cls_ids):
             b = float(self._bonus_array[c])
             if b <= 0.0:
@@ -527,25 +375,36 @@ class Miner:
                 continue
             idx = np.where(cm)[0]
             top = int(idx[int(np.argmax(scores[idx]))])
-            if scores[top] >= self._conf_thres_array[c] - b:
+            if scores[top] >= float(self._conf_thres_array[c]) - b:
                 keep[top] = True
         return keep
 
-    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]:
-        """Remove near-duplicate boxes across classes.
-
-        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.
-        """
+    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]:
+        """Vectorized per-class min-area / max-area / min-side / aspect filter."""
+        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.maximum(bw / np.maximum(bh, 1e-6), bh / np.maximum(bw, 1e-6))
+        keep = (
+            (bw >= self.min_side) & (bh >= self.min_side) &
+            (area >= self._min_box_area_array[cls_ids]) &
+            (area <= self.max_area_frac * image_area) &
+            (ar <= self.max_aspect_ratio)
+        )
+        return boxes[keep], scores[keep], cls_ids[keep]
+
+    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]:
+        """Suppress near-duplicate boxes ACROSS classes (same physical object
+        firing 2 classes, e.g. spray -> nozzle+track). Order by conf-margin then
+        area; keep highest, drop IoU>thresh others."""
         n = len(boxes)
         if n <= 1:
             return boxes, scores, cls_ids
@@ -576,141 +435,84 @@ class Miner:
         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,
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """Per-view post-processing: per-class NMS -> cap -> cross-class dedup."""
+    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]:
+        """sane filter -> top-k cap -> per-class NMS -> max_det cap -> 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
+        if len(scores) > self.pre_nms_topk:
+            top = np.argpartition(-scores, self.pre_nms_topk)[: self.pre_nms_topk]
+            boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]
         if len(boxes) > 1:
-            keep = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
+            keep = self._per_class_hard_nms(boxes, scores, cls_ids)
             boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
         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
-            )
+                boxes, scores, cls_ids, self.cross_iou_thresh)
         return boxes, scores, cls_ids
 
-    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:
-            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]
+    # ── Coordinate un-mapping (shared by both decode paths) ──────────────────
 
-        # 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]
+    def _unmap_and_finish(self, boxes: np.ndarray, scores: np.ndarray,
+                          cls_ids: np.ndarray, meta: dict
+                          ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Undo letterbox + square-pad, drop pad-center boxes, clip, post-process."""
+        pad_w, pad_h = meta["pad"]
+        ratio = meta["ratio"]
+        orig_w, orig_h = meta["orig_size"]
+        extra_left, extra_right = meta["extra_left"], meta["extra_right"]
 
-        if len(boxes) == 0:
-            return []
-
-        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
+        if extra_left:
+            boxes[:, [0, 2]] -= extra_left
+        if extra_left or extra_right:
+            cx = (boxes[:, 0] + boxes[:, 2]) * 0.5
+            inside = (cx >= 0) & (cx <= orig_w)
+            boxes, scores, cls_ids = boxes[inside], scores[inside], cls_ids[inside]
+            if len(boxes) == 0:
+                return boxes, scores, cls_ids
         boxes = self._clip_boxes(boxes, (orig_w, orig_h))
+        return self._per_view_pipeline(boxes, scores, cls_ids, (orig_w, orig_h))
 
-        # Box sanity filter (reduces FP)
-        boxes, scores, cls_ids = self._filter_sane_boxes(
-            boxes, scores, cls_ids, orig_size
-        )
-        if len(boxes) == 0:
-            return []
-
-        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)
-
-        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}")
+    # ── Decoding ─────────────────────────────────────────────────────────────
 
+    def _decode_final_dets(self, preds: np.ndarray, meta: dict
+                           ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """End2end output rows [x1, y1, x2, y2, conf, cls_id] (NMS in graph)."""
+        empty = (np.empty((0, 4), np.float32), np.empty(0, np.float32), np.empty(0, np.int32))
+        if preds.ndim == 3 and preds.shape[0] == 1:
+            preds = preds[0]
+        if preds.ndim != 2 or preds.shape[1] < 6:
+            raise ValueError(f"Unexpected final-det output shape: {preds.shape}")
+        boxes = preds[:, :4].astype(np.float32)
+        scores = preds[:, 4].astype(np.float32)
+        cls_ids = self.cls_remap[preds[:, 5].astype(np.int32)]
+        keep = self._conf_filter_mask(scores, cls_ids, meta["extra_left"])
+        boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
+        if len(boxes) == 0:
+            return empty
+        return self._unmap_and_finish(boxes, scores, cls_ids, meta)
+
+    def _decode_raw_yolo(self, preds: np.ndarray, meta: dict
+                         ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Raw YOLO output [1,C,N] or [1,N,C]; xywh + per-class scores."""
+        empty = (np.empty((0, 4), np.float32), np.empty(0, np.float32), np.empty(0, np.int32))
+        if preds.ndim != 3 or preds.shape[0] != 1:
+            raise ValueError(f"Unexpected raw output shape: {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}")
-
+            raise ValueError(f"Unexpected normalized raw 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)
@@ -718,196 +520,97 @@ class Miner:
             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]
+        keep = self._conf_filter_mask(scores, cls_ids, meta["extra_left"])
+        boxes_xywh, scores, cls_ids = boxes_xywh[keep], scores[keep], cls_ids[keep]
         if len(boxes_xywh) == 0:
-            return []
-
+            return empty
         boxes = self._xywh_to_xyxy(boxes_xywh)
+        return self._unmap_and_finish(boxes, scores, cls_ids, meta)
 
-        # 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))
+    def _decode(self, output: np.ndarray, meta: dict
+                ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        if output.ndim == 2 and output.shape[1] >= 6:
+            return self._decode_final_dets(output, meta)
+        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 6:
+            return self._decode_final_dets(output, meta)
+        return self._decode_raw_yolo(output, meta)
 
-        boxes, scores, cls_ids = self._filter_sane_boxes(
-            boxes, scores, cls_ids, (orig_w, orig_h)
-        )
-        if len(boxes) == 0:
-            return []
+    # ── Inference ────────────────────────────────────────────────────────────
 
-        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
+    def _predict_single_arrays(self, image: np.ndarray
+                               ) -> tuple[np.ndarray, np.ndarray, np.ndarray, dict]:
+        if not isinstance(image, np.ndarray) or image.ndim != 3 or image.shape[2] != 3:
+            raise ValueError(f"Expected HWC BGR image, got {type(image)} {getattr(image,'shape',None)}")
+        if image.dtype != np.uint8:
+            image = image.astype(np.uint8)
+        inp, meta = self._preprocess(image)
+        outputs = self.session.run(self.output_names, {self.input_name: inp})
+        boxes, scores, cls_ids = self._decode(outputs[0], meta)
+        return boxes, scores, cls_ids, meta
 
+    @staticmethod
+    def _to_boxes(boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray,
+                  orig_size: tuple[int, int]) -> list[BoundingBox]:
         results: list[BoundingBox] = []
+        orig_w, orig_h = orig_size
         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),
-                )
-            )
-
+            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)))),
+                cls_id=int(cls_id),
+                conf=float(max(0.0, min(1.0, conf))),
+            ))
         return results
 
-    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)
-
+    def _infer_single(self, image: np.ndarray) -> list[BoundingBox]:
+        boxes, scores, cls_ids, meta = self._predict_single_arrays(image)
+        if len(boxes) == 0:
+            return []
+        if self.use_cluster_boost and len(boxes) > 1:
+            scores = self._max_score_per_cluster(
+                boxes, cls_ids, boxes, scores, cls_ids, self._avg_iou)
+        return self._to_boxes(boxes, scores, cls_ids, meta["orig_size"])
+
+    def _infer_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        """Horizontal-flip TTA: union of original + flipped, per-class NMS,
+        same-class cluster-max boost, cross-class dedup."""
+        b0, s0, c0, meta = self._predict_single_arrays(image)
         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,
-            )
-            for b in boxes_flip
-        ]
-
-        all_boxes = boxes_orig + boxes_flip
-        if len(all_boxes) == 0:
+        bf, sf, cf, _ = self._predict_single_arrays(cv2.flip(image, 1))
+        if len(bf):
+            bf = bf.copy()
+            bf[:, [0, 2]] = w - bf[:, [2, 0]]  # mirror x back to original coords
+        coords = np.concatenate([b0, bf], axis=0) if len(bf) else b0
+        scores = np.concatenate([s0, sf], axis=0) if len(sf) else s0
+        cls_ids = np.concatenate([c0, cf], axis=0) if len(cf) else c0
+        if len(coords) == 0:
             return []
-
-        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)
-
-        hard_keep = self._per_class_hard_nms(coords, scores, cls_ids, self.iou_thres)
-        if len(hard_keep) == 0:
+        keep = self._per_class_hard_nms(coords, scores, cls_ids)
+        if len(keep) == 0:
             return []
-        if len(hard_keep) > self.max_det:
-            top = np.argsort(-scores[hard_keep])[: self.max_det]
-            hard_keep = hard_keep[top]
-
-        # Class-aware cluster-max score boost (fixes the silent cross-class
-        # leak in the previous _max_score_per_cluster).
+        if len(keep) > self.max_det:
+            keep = keep[np.argsort(-scores[keep])[: self.max_det]]
         boosted = self._max_score_per_cluster(
-            coords[hard_keep], cls_ids[hard_keep],
-            coords, scores, cls_ids, self.iou_thres,
-        )
-
-        kept_coords = coords[hard_keep]
-        kept_cls = cls_ids[hard_keep]
-        if len(kept_coords) > 1:
-            kept_coords, boosted, kept_cls = self._cross_class_dedup_op(
-                kept_coords, boosted, kept_cls, self.cross_iou_thresh
-            )
-
-        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]),
-            )
-            for j in range(len(kept_coords))
-        ]
-
-    def predict_batch(
-        self,
-        batch_images: list[ndarray],
-        offset: int,
-        n_keypoints: int,
-    ) -> list[TVFrameResult]:
+            coords[keep], cls_ids[keep], coords, scores, cls_ids, self._avg_iou)
+        kb, kc = coords[keep], cls_ids[keep]
+        if len(kb) > 1:
+            kb, boosted, kc = self._cross_class_dedup_op(kb, boosted, kc, self.cross_iou_thresh)
+        return self._to_boxes(kb, boosted, kc, meta["orig_size"])
+
+    def predict_batch(self, batch_images: list[ndarray], offset: int,
+                      n_keypoints: int) -> list[TVFrameResult]:
         results: list[TVFrameResult] = []
-
-        for frame_number_in_batch, image in enumerate(batch_images):
+        for idx, image in enumerate(batch_images):
             try:
-                if self.use_tta:
-                    boxes = self._predict_tta(image)
-                else:
-                    boxes = self._predict_single(image)
+                boxes = self._infer_tta(image) if self.use_tta else self._infer_single(image)
             except Exception as e:
-                print(f"⚠️ Inference failed for frame {offset + frame_number_in_batch}: {e}")
+                print(f"Inference failed for frame {offset + idx}: {e}")
                 boxes = []
-
-            results.append(
-                TVFrameResult(
-                    frame_id=offset + frame_number_in_batch,
-                    boxes=boxes,
-                    keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
-                )
-            )
-
-        return results
+            keypoints = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
+            results.append(TVFrameResult(frame_id=offset + idx, boxes=boxes, keypoints=keypoints))
+        return results

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:756f81aa467d482bb6508d172cc20c34fbb4cb7f15eeabb5538b858a1e561c94
-size 9824686
+oid sha256:d5b41a6a181550440eda18f4af1f7a70e2e1edfa508fd5c29959e63ce824daeb
+size 9842895