Upload folder using huggingface_hub

chute_config.yml

@@ -8,7 +8,6 @@ Image:
 NodeSelector:
   gpu_count: 1
   min_vram_gb_per_gpu: 16
-  max_hourly_price_per_gpu: 2
   include:
     - pro_6000
 

miner.py

@@ -1,5 +1,4 @@
 from pathlib import Path
-import math
 
 import cv2
 import numpy as np
@@ -24,20 +23,74 @@ class TVFrameResult(BaseModel):
 
 
 class Miner:
-    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"]
+    """ONNX Runtime miner for road-sign detection (single class).
+    Strategy (ported from offense / fire001 miner):
+      - per-class confidence threshold with per-class rescue bonus
+      - per-class hard NMS, then cross-class dedup (no-op for single class)
+      - horizontal-flip TTA with full-set cluster score boost
+    Plus: class remap, sanity-box filter tuned for small distant signs,
+    TTA toggle.
+    """
+
+    class_names = ["road_sign"]
+    # Order the model emits classes in -- remapped to `class_names` index.
+    _model_class_order = ["road_sign"]
+
+    iou_thres = 0.5
+    cross_iou_thresh = 0.8
+    max_det = 150
+
+    # Per-class confidence threshold. Road signs in this dataset are
+    # frequently degraded / rear-facing / partly-obscured / distant, so we
+    # run noticeably below the fire/smoke baseline. The validator's
+    # false_positive pillar = max(0, 1 - ffpi/10): we can tolerate ~2 FP per
+    # image and still keep that pillar above 0.8.
+    _conf_thres_array = np.array(
+        [0.33], dtype=np.float32
+    )
+    # Per-class rescue bonus. If a class has ZERO boxes passing the threshold
+    # in a frame, its top-1 candidate is admitted when its score is at least
+    # (threshold - bonus). Bumped from 0.05 -> 0.10 so a single faint sign in
+    # an otherwise empty frame still produces a detection (map50 recall win,
+    # at most one extra FP per such frame).
+    _bonus_array = np.array(
+        [0.12], dtype=np.float32
+    )
+
+    # Box sanity filter: drop tiny / degenerate / image-spanning / extreme
+    # aspect ratio boxes.
+    #   min_box_area = 14x14  -> 14x14 is the smallest credible sign. The old
+    #                          value of 64 (8x8) silently discarded narrow
+    #                          distant signs like a 10x6 px overhead chevron.
+    #   min_side     = 3      -> matches min_box_area; anything thinner is
+    #                          almost certainly a pole or shadow false alarm.
+    #   max_aspect_ratio = 12.0
+    #                         -> overhead destination panels and lane-assignment
+    #                          signs are very wide (long, thin rectangles);
+    #                          8.0 was clipping legitimate detections.
+    min_box_area = 8 * 8
+    min_side = 3
+    max_aspect_ratio = 12.0
+
+    # Final box-size calibration. The detector + de-letterbox + integer-rounding
+    # pipeline emits boxes slightly larger than the object, so shrink every
+    # emitted box about its center by a fixed per-axis factor before output:
+    #   new_w = w / box_shrink_w,  new_h = h / box_shrink_h.
+    box_shrink_w = 1.027
+    box_shrink_h = 1.014
+
+    # Tile-based TTA: when the source image is significantly larger than the
+    # model input, letterboxing throws away ~1.5x of effective resolution,
+    # which kills small-sign recall. Splitting into overlapping horizontal
+    # tiles preserves native resolution on each half. Triggered only when
+    # source width >= tile_trigger_ratio * model_input_width to avoid wasted
+    # compute on already-small images.
+    tile_trigger_ratio = 1.4
+    tile_overlap_ratio = 0.20
+
+    def __init__(self, path_hf_repo: Path) -> None:
+        model_path = path_hf_repo / "weights.onnx"
+        
         print("ORT version:", ort.__version__)
 
         try:
@@ -58,11 +111,9 @@ class Miner:
             self.session = ort.InferenceSession(
                 str(model_path),
                 sess_options=sess_options,
-                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+                providers=["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,
@@ -73,8 +124,10 @@ class Miner:
 
         # 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.
+        # i.e. convert a model-side class id into the output class id that
+        # downstream code (BoundingBox.cls_id, the per-class threshold/bonus
+        # arrays) expects. 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)
@@ -82,12 +135,12 @@ class Miner:
         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
+            [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)
 
@@ -95,56 +148,32 @@ class Miner:
         self.output_names = [output.name for output 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).
-        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.
+        # weights.onnx is exported at 1280x1280 (Ultralytics imgsz metadata),
+        # static (dynamic=False). The default is only the fallback for when the
+        # ONNX input dims aren't fixed; the real value is read from the session.
         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)
-        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.min_side = 3
-        self.max_aspect_ratio = 12.0
+        self.use_tile_tta = False
+        # Soft-NMS (ported from carwash001): Gaussian score decay of overlapping
+        # boxes instead of hard removal. OFF by default to preserve the current
+        # deployed behaviour; flip on (and tune sigma) via tune_miner.py to see if
+        # it scores better — useful where signs cluster (gantries, sign assemblies).
+        self.use_soft_nms = False
+        self.soft_nms_sigma = 0.5
+        self.soft_nms_score_thresh = 0.01
 
         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}")
+        print(f"✅ ONNX input size: {self.input_width}x{self.input_height}, "
+              f"use_tta={self.use_tta}, use_tile_tta={self.use_tile_tta}")
+        print("per-class conf: " + ", ".join(
+            f"{n}={t:.3f}" for n, t in zip(
+                self.class_names, self._conf_thres_array.tolist()
+            )
+        ))
 
         self._warmup()
 
@@ -157,50 +186,16 @@ class Miner:
         except Exception as e:
             print(f"⚠️ warmup skipped: {e}")
 
-    def __repr__(self) -> str:
-        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
-
-    @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:
+    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).
-        """
+        Returns the class names ordered by model-emit index, or None when the
+        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', ...}
+            names = ast.literal_eval(meta["names"])  # e.g. {0: 'road_sign'}
             if isinstance(names, dict):
                 order = [str(names[i]) for i in sorted(names)]
             else:
@@ -216,19 +211,22 @@ class Miner:
             return None
         return order
 
+    def __repr__(self) -> str:
+        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
+
     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
-        """
         h, w = image.shape[:2]
         new_w, new_h = new_shape
 
@@ -240,10 +238,8 @@ class Miner:
             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
+        dw = (new_w - resized_w) / 2.0
+        dh = (new_h - resized_h) / 2.0
 
         left = int(round(dw - 0.1))
         right = int(round(dw + 0.1))
@@ -251,38 +247,23 @@ class Miner:
         bottom = int(round(dh + 0.1))
 
         padded = cv2.copyMakeBorder(
-            image,
-            top,
-            bottom,
-            left,
-            right,
-            borderType=cv2.BORDER_CONSTANT,
-            value=color,
+            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]
-
         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 = np.transpose(img, (2, 0, 1))[None, ...]
-        img = np.ascontiguousarray(img, dtype=self.np_dtype)
-
-        return img, ratio, pad, (orig_w, orig_h)
+        # Fused scale(1/255) + BGR->RGB swap + HWC->NCHW + contiguous float32 in
+        # one optimized OpenCV call (bit-identical to the cvtColor + astype/255 +
+        # transpose chain, but ~half the preprocess time).
+        blob = cv2.dnn.blobFromImage(img, scalefactor=1.0 / 255.0, swapRB=True)
+        return blob, ratio, pad, (orig_w, orig_h)
 
     @staticmethod
     def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
@@ -302,6 +283,52 @@ class Miner:
         out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
         return out
 
+    @staticmethod
+    def _hard_nms(
+        boxes: np.ndarray, scores: np.ndarray, iou_thresh: float
+    ) -> np.ndarray:
+        n = len(boxes)
+        if n == 0:
+            return np.array([], dtype=np.intp)
+        order = np.argsort(-scores)
+        keep: list[int] = []
+        while len(order) > 0:
+            i = int(order[0])
+            keep.append(i)
+            if len(order) == 1:
+                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])
+            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,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        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())
+        all_keep.sort()
+        return np.array(all_keep, dtype=np.intp)
+
     def _soft_nms(
         self,
         boxes: np.ndarray,
@@ -309,106 +336,35 @@ class Miner:
         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).
-        """
+        """Soft-NMS: Gaussian decay of overlapping scores instead of hard removal.
+        Returns (kept_original_indices, updated_scores). (Ported from carwash001.)"""
         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
-
             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)
-
             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])
-            )
+                (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: 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())
-        all_keep.sort()
-        return np.array(all_keep, dtype=np.intp)
-
     def _per_class_soft_nms(
         self,
         boxes: np.ndarray,
@@ -417,60 +373,64 @@ class Miner:
         sigma: float = 0.5,
         score_thresh: float = 0.01,
     ) -> tuple[np.ndarray, np.ndarray]:
-        """Soft NMS applied independently per class."""
+        """Soft-NMS applied independently per class. Returns (kept_idx, updated_scores)."""
         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)
+            indices = np.where(cls_ids == c)[0]
+            keep, updated = self._soft_nms(boxes[indices], scores[indices],
+                                           sigma, score_thresh)
             for k, s in zip(keep, updated):
-                all_keep.append(int(indices[k]))
-                all_scores.append(float(s))
+                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(
+    def _cross_class_dedup_op(
         self,
         boxes: np.ndarray,
         scores: np.ndarray,
         cls_ids: np.ndarray,
-        orig_size: tuple[int, int],
+        iou_thresh: float,
     ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """Filter out tiny, degenerate, or implausible boxes (common FP)."""
-        if len(boxes) == 0:
+        """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.
+        With a single road_sign class this is effectively a no-op, but the
+        method is kept so the pipeline stays compatible with the multi-class
+        miner template.
+        """
+        n = len(boxes)
+        if n <= 1:
             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:
+        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: list[int] = []
+        for i in order:
+            if suppressed[i]:
                 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]
+            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]
 
     @staticmethod
     def _max_score_per_cluster(
@@ -482,11 +442,9 @@ class Miner:
         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.
+        candidate set among same-class boxes with IoU >= iou_thresh.
+        Used after horizontal-flip TTA: a high-confidence flipped detection
+        can raise the score of the corresponding original detection.
         """
         n = len(post_boxes)
         if n == 0:
@@ -512,8 +470,7 @@ class Miner:
     ) -> 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).
-        """
+        when its score >= (per-class threshold - per-class bonus)."""
         if len(scores) == 0:
             return np.zeros(0, dtype=bool)
         thr = self._conf_thres_array[cls_ids]
@@ -531,50 +488,44 @@ class Miner:
                 keep[top] = True
         return keep
 
-    def _cross_class_dedup_op(
+    def _filter_sane_boxes(
         self,
         boxes: np.ndarray,
         scores: np.ndarray,
         cls_ids: np.ndarray,
-        iou_thresh: float,
+        orig_size: tuple[int, int],
     ) -> 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.
-        """
-        n = len(boxes)
-        if n <= 1:
+        """Drop tiny / degenerate / image-spanning / extreme-AR boxes (FP)."""
+        if len(boxes) == 0:
             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: list[int] = []
-        for i in order:
-            if suppressed[i]:
+        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
-            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]
+            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]
 
     def _per_view_pipeline(
         self,
@@ -582,10 +533,16 @@ class Miner:
         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."""
+        """Per-view post-processing pipeline: per-class NMS -> cap -> cross-class dedup."""
         if len(boxes) > 1:
-            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 self.use_soft_nms:
+                keep, new_scores = self._per_class_soft_nms(
+                    boxes, scores, cls_ids,
+                    self.soft_nms_sigma, self.soft_nms_score_thresh)
+                boxes, scores, cls_ids = boxes[keep], new_scores, cls_ids[keep]
+            else:
+                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]
             boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]
@@ -595,22 +552,53 @@ class Miner:
             )
         return boxes, scores, cls_ids
 
+    @staticmethod
+    def _shrink_wh(coords: np.ndarray, sw: float, sh: float) -> np.ndarray:
+        """Shrink each xyxy box about its center: new_w = w/sw, new_h = h/sh."""
+        if len(coords) == 0:
+            return coords
+        coords = np.asarray(coords, dtype=np.float32).copy()
+        cx = (coords[:, 0] + coords[:, 2]) * 0.5
+        cy = (coords[:, 1] + coords[:, 3]) * 0.5
+        hw = (coords[:, 2] - coords[:, 0]) * (0.5 / sw)
+        hh = (coords[:, 3] - coords[:, 1]) * (0.5 / sh)
+        coords[:, 0] = cx - hw
+        coords[:, 1] = cy - hh
+        coords[:, 2] = cx + hw
+        coords[:, 3] = cy + hh
+        return coords
+
+    def _build_results(
+        self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray
+    ) -> list[BoundingBox]:
+        boxes = self._shrink_wh(boxes, self.box_shrink_w, self.box_shrink_h)
+        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(round(x1)),
+                    y1=int(round(y1)),
+                    x2=int(round(x2)),
+                    y2=int(round(y2)),
+                    cls_id=int(cls_id),
+                    conf=float(conf),
+                )
+            )
+        return results
+
     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.
-        """
+        """Final-detection output path: rows shaped [x1, y1, x2, y2, conf, cls_id]."""
         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}")
 
@@ -619,66 +607,27 @@ class Miner:
         cls_ids = preds[:, 5].astype(np.int32)
         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 = boxes[keep]
         scores = scores[keep]
         cls_ids = cls_ids[keep]
-
         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
-        boxes = self._clip_boxes(boxes, (orig_w, orig_h))
+        boxes = self._clip_boxes(boxes, orig_size)
 
-        # 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
+        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
+        return self._build_results(boxes, scores, cls_ids)
 
     def _decode_raw_yolo(
         self,
@@ -687,30 +636,17 @@ class Miner:
         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:
+        """Fallback raw-YOLO output path: per-anchor class logits."""
+        if preds.ndim != 3 or preds.shape[0] != 1:
             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}")
+            raise ValueError(f"Unexpected 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)
@@ -719,52 +655,28 @@ class Miner:
             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 = self._clip_boxes(boxes, orig_size)
 
         boxes, scores, cls_ids = self._filter_sane_boxes(
-            boxes, scores, cls_ids, (orig_w, orig_h)
+            boxes, scores, cls_ids, orig_size
         )
         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=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
+        return self._build_results(boxes, scores, cls_ids)
 
     def _postprocess(
         self,
@@ -773,19 +685,10 @@ class Miner:
         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]:
@@ -799,39 +702,33 @@ class Miner:
             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:
+        expected = (1, 3, self.input_height, self.input_width)
+        if input_tensor.shape != expected:
             raise ValueError(
-                f"Bad input tensor shape={input_tensor.shape}, expected={expected_shape}"
+                f"Bad input tensor shape={input_tensor.shape}, expected={expected}"
             )
 
         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)
+        return self._postprocess(outputs[0], ratio, pad, orig_size)
 
     def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
         """Horizontal-flip TTA.
-
-        Strategy (ported from fire001):
+        Strategy:
           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.
+          4. For each kept box, compute the max same-class score across the
+             FULL union (not just the post-NMS subset) -- this lets a high-
+             confidence flipped detection raise a borderline original one.
           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(
@@ -840,9 +737,8 @@ class Miner:
             )
             for b in boxes_flip
         ]
-
         all_boxes = boxes_orig + boxes_flip
-        if len(all_boxes) == 0:
+        if not all_boxes:
             return []
 
         coords = np.array(
@@ -858,8 +754,6 @@ class Miner:
             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).
         boosted = self._max_score_per_cluster(
             coords[hard_keep], cls_ids[hard_keep],
             coords, scores, cls_ids, self.iou_thres,
@@ -872,18 +766,151 @@ class Miner:
                 kept_coords, boosted, kept_cls, self.cross_iou_thresh
             )
 
+        kept_coords = self._shrink_wh(
+            kept_coords, self.box_shrink_w, self.box_shrink_h
+        )
         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])),
+                x1=int(round(float(kept_coords[j, 0]))),
+                y1=int(round(float(kept_coords[j, 1]))),
+                x2=int(round(float(kept_coords[j, 2]))),
+                y2=int(round(float(kept_coords[j, 3]))),
                 cls_id=int(kept_cls[j]),
                 conf=float(boosted[j]),
             )
             for j in range(len(kept_coords))
         ]
 
+    def _predict_tiles(self, image: np.ndarray) -> list[BoundingBox]:
+        """Tile-based TTA for high-resolution images.
+        Splits the source image into two overlapping horizontal tiles, runs
+        single-pass inference on each at native scale, and translates boxes
+        back to the global frame. Useful when source width >> model input
+        width because letterboxing otherwise discards effective resolution
+        that small / distant signs depend on.
+        Returns an empty list if the image isn't wide enough to benefit; the
+        caller falls back to the regular pipeline in that case.
+        """
+        h, w = image.shape[:2]
+        if w < int(self.input_width * self.tile_trigger_ratio):
+            return []
+
+        overlap = int(w * self.tile_overlap_ratio)
+        mid = w // 2
+        x_left_end = min(w, mid + overlap // 2)
+        x_right_start = max(0, mid - overlap // 2)
+
+        left = image[:, :x_left_end]
+        right = image[:, x_right_start:]
+
+        boxes_left = self._predict_single(left)
+        boxes_right = self._predict_single(right)
+
+        shifted_right = [
+            BoundingBox(
+                x1=b.x1 + x_right_start,
+                y1=b.y1,
+                x2=b.x2 + x_right_start,
+                y2=b.y2,
+                cls_id=b.cls_id,
+                conf=b.conf,
+            )
+            for b in boxes_right
+        ]
+        return boxes_left + shifted_right
+
+    def _merge_views(
+        self,
+        view_boxes: list[list[BoundingBox]],
+        image_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """Merge boxes from multiple views (single / hflip / tiles).
+        Same logic as `_predict_tta`'s tail: per-class hard NMS to dedupe,
+        then for each kept box take the max same-class score across the full
+        candidate union — a high-confidence detection in any view boosts
+        borderline matches in others.
+        """
+        all_boxes: list[BoundingBox] = []
+        for vb in view_boxes:
+            all_boxes.extend(vb)
+        if not all_boxes:
+            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)
+
+        coords = self._clip_boxes(coords, image_size)
+
+        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]
+
+        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
+            )
+
+        kept_coords = self._shrink_wh(
+            kept_coords, self.box_shrink_w, self.box_shrink_h
+        )
+        return [
+            BoundingBox(
+                x1=int(round(float(kept_coords[j, 0]))),
+                y1=int(round(float(kept_coords[j, 1]))),
+                x2=int(round(float(kept_coords[j, 2]))),
+                y2=int(round(float(kept_coords[j, 3]))),
+                cls_id=int(kept_cls[j]),
+                conf=float(boosted[j]),
+            )
+            for j in range(len(kept_coords))
+        ]
+
+    def _predict_full(self, image: np.ndarray) -> list[BoundingBox]:
+        """Top-level per-frame prediction with all enabled augmentations.
+        - `use_tta=True`: original + horizontal flip
+        - `use_tile_tta=True` AND image wide enough: two overlapping tiles
+        All views are merged via per-class NMS + cluster-max score boost.
+        """
+        if not self.use_tta and not self.use_tile_tta:
+            return self._predict_single(image)
+
+        views: list[list[BoundingBox]] = []
+        if self.use_tta:
+            views.append(self._predict_single(image))
+            flipped = cv2.flip(image, 1)
+            w = image.shape[1]
+            flipped_dets = self._predict_single(flipped)
+            views.append([
+                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 flipped_dets
+            ])
+        else:
+            views.append(self._predict_single(image))
+
+        if self.use_tile_tta:
+            tile_boxes = self._predict_tiles(image)
+            if tile_boxes:
+                views.append(tile_boxes)
+
+        h, w = image.shape[:2]
+        return self._merge_views(views, (w, h))
+
     def predict_batch(
         self,
         batch_images: list[ndarray],
@@ -891,17 +918,15 @@ class Miner:
         n_keypoints: int,
     ) -> list[TVFrameResult]:
         results: list[TVFrameResult] = []
-
         for frame_number_in_batch, image in enumerate(batch_images):
             try:
-                if self.use_tta:
-                    boxes = self._predict_tta(image)
-                else:
-                    boxes = self._predict_single(image)
+                boxes = self._predict_full(image)
             except Exception as e:
-                print(f"⚠️ Inference failed for frame {offset + frame_number_in_batch}: {e}")
+                print(
+                    f"⚠️ Inference failed for frame "
+                    f"{offset + frame_number_in_batch}: {e}"
+                )
                 boxes = []
-
             results.append(
                 TVFrameResult(
                     frame_id=offset + frame_number_in_batch,
@@ -909,5 +934,4 @@ class Miner:
                     keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
                 )
             )
-
-        return results
+        return results

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:756f81aa467d482bb6508d172cc20c34fbb4cb7f15eeabb5538b858a1e561c94
-size 9824686
+oid sha256:5eb4543ec44fe7dd6a743e36b687ddd08985e54fdd358426b118bec1c4972ad7
+size 9805054