scorevision: push artifact

miner.py

@@ -0,0 +1,806 @@
+from pathlib import Path
+import math
+
+import cv2
+import numpy as np
+import onnxruntime as ort
+from numpy import ndarray
+from pydantic import BaseModel
+
+
+class BoundingBox(BaseModel):
+    x1: int
+    y1: int
+    x2: int
+    y2: int
+    cls_id: int
+    conf: float
+
+
+class TVFrameResult(BaseModel):
+    frame_id: int
+    boxes: list[BoundingBox]
+    keypoints: list[tuple[int, int]]
+
+
+class Miner:
+    """ONNX Runtime miner for fire / smoke / fire_extinguisher detection.
+
+    Strategy (ported from offense miner):
+      - per-class confidence threshold with per-class rescue bonus
+      - per-class hard NMS, then cross-class dedup
+      - horizontal-flip TTA with full-set cluster score boost
+    Plus fire001 specifics: class remap, sanity-box filter, TTA toggle.
+    """
+
+    class_names = ["fire", "smoke", "fire extinguisher"]
+    _cls_fire = 0  # index in class_names
+    _cls_smoke = 1  # index in class_names
+    _cls_fire_extinguisher = 2  # index in class_names
+    _nested_zone_classes = (_cls_fire, _cls_smoke)
+    # Order the model emits classes in -- remapped to `class_names` index.
+    _model_class_order = ["fire", "fire extinguisher", "smoke"]
+
+    iou_thres = 0.55
+    cross_iou_thresh = 0.8
+    max_det = 150
+    nested_contain_ratio = 0.95
+    #"fire", "smoke", "fire extinguisher"
+    _conf_thres_array = np.array([0.22, 0.3, 0.42], dtype=np.float32)
+    _bonus_array = np.array([0.18, 0.27, 0.395], dtype=np.float32)
+    # Box sanity filter (fire001-specific FP reduction): drop tiny / degenerate
+    # / image-spanning / extreme aspect ratio boxes.
+    min_box_area = 14 * 14
+    min_side = 8
+    max_aspect_ratio = 8.0
+
+    def __init__(self, path_hf_repo: Path) -> None:
+        model_path = path_hf_repo / "weights.onnx"
+        self.cls_remap = np.array(
+            [self.class_names.index(n) for n in self._model_class_order],
+            dtype=np.int32,
+        )
+        print("ORT version:", ort.__version__)
+
+        try:
+            ort.preload_dlls()
+            print("✅ onnxruntime.preload_dlls() success")
+        except Exception as 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
+
+        try:
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+            )
+            print("✅ Created ORT session with preferred CUDA provider list")
+        except Exception as e:
+            print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}")
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )
+
+        print("ORT session providers:", self.session.get_providers())
+
+        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.input_shape = self.session.get_inputs()[0].shape
+
+        self.input_height = self._safe_dim(self.input_shape[2], default=1280)
+        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
+
+        self.use_tta = True
+
+        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("per-class conf: " + ", ".join(
+            f"{n}={t:.3f}" for n, t in zip(
+                self.class_names, self._conf_thres_array.tolist()
+            )
+        ))
+
+    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]]:
+        h, w = image.shape[:2]
+        new_w, new_h = new_shape
+
+        ratio = min(new_w / w, new_h / h)
+        resized_w = int(round(w * ratio))
+        resized_h = int(round(h * ratio))
+
+        if (resized_w, resized_h) != (w, h):
+            interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR
+            image = cv2.resize(image, (resized_w, resized_h), interpolation=interp)
+
+        dw = (new_w - resized_w) / 2.0
+        dh = (new_h - resized_h) / 2.0
+
+        left = int(round(dw - 0.1))
+        right = int(round(dw + 0.1))
+        top = int(round(dh - 0.1))
+        bottom = int(round(dh + 0.1))
+
+        padded = cv2.copyMakeBorder(
+            image, top, bottom, left, right,
+            borderType=cv2.BORDER_CONSTANT, value=color,
+        )
+        return padded, ratio, (dw, dh)
+
+    def _preprocess(
+        self, image: ndarray
+    ) -> tuple[np.ndarray, float, tuple[float, float], tuple[int, int]]:
+        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=np.float32)
+        return img, ratio, pad, (orig_w, orig_h)
+
+    @staticmethod
+    def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
+        w, h = image_size
+        boxes[:, 0] = np.clip(boxes[:, 0], 0, w - 1)
+        boxes[:, 1] = np.clip(boxes[:, 1], 0, h - 1)
+        boxes[:, 2] = np.clip(boxes[:, 2], 0, w - 1)
+        boxes[:, 3] = np.clip(boxes[:, 3], 0, h - 1)
+        return boxes
+
+    @staticmethod
+    def _xywh_to_xyxy(boxes: np.ndarray) -> np.ndarray:
+        out = np.empty_like(boxes)
+        out[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.0
+        out[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.0
+        out[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.0
+        out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
+        return out
+
+    @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)
+
+    @staticmethod
+    def _box_mostly_contained(
+        outer: np.ndarray, inner: np.ndarray, ratio: float
+    ) -> bool:
+        """True when at least `ratio` of inner's area lies inside outer."""
+        xx1 = max(float(outer[0]), float(inner[0]))
+        yy1 = max(float(outer[1]), float(inner[1]))
+        xx2 = min(float(outer[2]), float(inner[2]))
+        yy2 = min(float(outer[3]), float(inner[3]))
+        inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1)
+        inner_area = max(
+            1e-7,
+            (float(inner[2]) - float(inner[0])) * (float(inner[3]) - float(inner[1])),
+        )
+        return inter / inner_area >= ratio
+
+    def _nested_zone_filter(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Among nested fire/smoke pairs (>=95% containment), keep higher conf."""
+        n = len(boxes)
+        if n <= 1:
+            return boxes, scores, cls_ids
+
+        ratio = self.nested_contain_ratio
+        boxes = np.asarray(boxes, dtype=np.float32)
+        scores = np.asarray(scores, dtype=np.float32)
+        cls_ids = np.asarray(cls_ids, dtype=np.int32)
+        suppress = np.zeros(n, dtype=bool)
+        for cls_id in self._nested_zone_classes:
+            class_idx = np.where(cls_ids == cls_id)[0]
+            if len(class_idx) <= 1:
+                continue
+            for a in range(len(class_idx)):
+                i = int(class_idx[a])
+                if suppress[i]:
+                    continue
+                bi = boxes[i]
+                for b in range(a + 1, len(class_idx)):
+                    j = int(class_idx[b])
+                    if suppress[j]:
+                        continue
+                    bj = boxes[j]
+                    nested = (
+                        self._box_mostly_contained(bi, bj, ratio)
+                        or self._box_mostly_contained(bj, bi, ratio)
+                    )
+                    if not nested:
+                        continue
+                    if scores[i] >= scores[j]:
+                        suppress[j] = True
+                    else:
+                        suppress[i] = True
+                        break
+
+        keep = ~suppress
+        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]:
+        """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.
+        This suppresses the case where the same physical object is detected
+        as multiple classes (e.g. fire vs smoke on the same flames).
+
+        Fire extinguisher is exempt: it is a distinct object and may overlap
+        fire/smoke boxes in scene without being a duplicate detection.
+        """
+        n = len(boxes)
+        if n <= 1:
+            return boxes, scores, cls_ids
+        boxes = np.asarray(boxes, dtype=np.float32)
+        scores = np.asarray(scores, dtype=np.float32)
+        cls_ids = np.asarray(cls_ids, dtype=np.int32)
+        ext_cls = self._cls_fire_extinguisher
+        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(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
+            # Never cross-suppress fire extinguisher vs fire/smoke.
+            dup &= ~((cls_ids == ext_cls) | (cls_ids[i] == ext_cls))
+            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(
+        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.
+
+        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:
+            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 _loose_conf_mask(
+        self, scores: np.ndarray, cls_ids: np.ndarray
+    ) -> np.ndarray:
+        """Pre-filter: keep candidates that could pass threshold or bonus rescue."""
+        floor = self._conf_thres_array[cls_ids] - self._bonus_array[cls_ids]
+        return scores >= floor
+
+    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 -- admit top-1 when score >= (threshold - bonus). Runs after
+        sane-box filtering so tiny FPs cannot block bonus rescue."""
+        if len(scores) == 0:
+            return np.zeros(0, dtype=bool)
+        thr = self._conf_thres_array[cls_ids]
+        keep = scores >= thr
+        ext_cls = self._cls_fire_extinguisher
+        for c in np.unique(cls_ids):
+            b = float(self._bonus_array[c])
+            if b <= 0.0:
+                continue
+            cm = cls_ids == c
+            idx = np.where(cm)[0]
+            top = int(idx[int(np.argmax(scores[idx]))])
+            floor = float(self._conf_thres_array[c] - b)
+            if scores[top] < floor:
+                continue
+            if c == ext_cls:
+                keep[top] = True
+            elif not keep[cm].any():
+                keep[top] = True
+        return keep
+
+    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]:
+        """Drop tiny / degenerate / image-spanning / extreme-AR boxes (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]
+
+    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 pipeline: per-class NMS -> nested filter -> cap -> cross-class dedup -> smoke-merge."""
+        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]
+        boxes, scores, cls_ids = self._nested_zone_filter(
+            boxes, scores, cls_ids
+        )
+        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._smoke_merge(boxes, scores, cls_ids)
+        return boxes, scores, cls_ids
+
+    def _smoke_merge(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        iou_thresh: float = 0.05,
+        center_dist_frac: float = 0.15,
+    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Union-find merge of smoke boxes that overlap (IoU>0.05) OR have
+        centers within 15% of image diagonal. Each cluster collapses to its
+        bounding hull with max conf. Non-smoke classes pass through.
+        """
+        if len(boxes) == 0:
+            return boxes, scores, cls_ids
+        smoke_mask = cls_ids == self._cls_smoke
+        if smoke_mask.sum() <= 1:
+            return boxes, scores, cls_ids
+        smoke_idx = np.where(smoke_mask)[0]
+        non_smoke_idx = np.where(~smoke_mask)[0]
+        s_boxes = boxes[smoke_idx]
+        s_scores = scores[smoke_idx]
+        n = len(s_boxes)
+        parent = list(range(n))
+        def find(i: int) -> int:
+            while parent[i] != i:
+                parent[i] = parent[parent[i]]
+                i = parent[i]
+            return i
+        def union(i: int, j: int) -> None:
+            a, b = find(i), find(j)
+            if a != b:
+                parent[b] = a
+        h_img = float(self.input_h)
+        w_img = float(self.input_w)
+        diag = (w_img ** 2 + h_img ** 2) ** 0.5
+        for i in range(n):
+            for j in range(i + 1, n):
+                a = s_boxes[i]; b = s_boxes[j]
+                ix1, iy1 = max(a[0], b[0]), max(a[1], b[1])
+                ix2, iy2 = min(a[2], b[2]), min(a[3], b[3])
+                iw = max(0.0, float(ix2 - ix1)); ih = max(0.0, float(iy2 - iy1))
+                inter = iw * ih
+                ua = float((a[2]-a[0])*(a[3]-a[1])) + float((b[2]-b[0])*(b[3]-b[1])) - inter
+                iou = inter / ua if ua > 0 else 0.0
+                ax = (a[0]+a[2])/2; ay = (a[1]+a[3])/2
+                bx = (b[0]+b[2])/2; by = (b[1]+b[3])/2
+                cd = float(((ax-bx)**2 + (ay-by)**2) ** 0.5) / diag if diag > 0 else 0.0
+                if iou > iou_thresh or cd < center_dist_frac:
+                    union(i, j)
+        clusters: dict[int, list[int]] = {}
+        for i in range(n):
+            clusters.setdefault(find(i), []).append(i)
+        merged_boxes_list = []
+        merged_scores_list = []
+        for grp in clusters.values():
+            arr = s_boxes[grp]
+            xs1 = float(np.min(arr[:, 0])); ys1 = float(np.min(arr[:, 1]))
+            xs2 = float(np.max(arr[:, 2])); ys2 = float(np.max(arr[:, 3]))
+            merged_boxes_list.append([xs1, ys1, xs2, ys2])
+            merged_scores_list.append(float(np.max(s_scores[grp])))
+        merged_boxes_arr = np.array(merged_boxes_list, dtype=boxes.dtype)
+        merged_scores_arr = np.array(merged_scores_list, dtype=scores.dtype)
+        merged_cls_arr = np.full(len(merged_boxes_arr), self._cls_smoke, dtype=cls_ids.dtype)
+        if len(non_smoke_idx):
+            out_boxes = np.concatenate([boxes[non_smoke_idx], merged_boxes_arr], axis=0)
+            out_scores = np.concatenate([scores[non_smoke_idx], merged_scores_arr], axis=0)
+            out_cls = np.concatenate([cls_ids[non_smoke_idx], merged_cls_arr], axis=0)
+        else:
+            out_boxes = merged_boxes_arr
+            out_scores = merged_scores_arr
+            out_cls = merged_cls_arr
+        return out_boxes, out_scores, out_cls
+
+    @staticmethod
+    def _build_results(
+        boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray
+    ) -> list[BoundingBox]:
+        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_final_dets(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """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}")
+
+        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]
+
+        keep = self._loose_conf_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
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+
+        boxes, scores, cls_ids = self._filter_sane_boxes(
+            boxes, scores, cls_ids, orig_size
+        )
+        if len(boxes) == 0:
+            return []
+
+        keep = self._conf_filter_mask(scores, cls_ids)
+        boxes = boxes[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+        if len(boxes) == 0:
+            return []
+
+        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,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """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}")
+        preds = preds[0]
+        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 raw output shape: {preds.shape}")
+
+        boxes_xywh = preds[:, :4].astype(np.float32)
+        cls_part = preds[:, 4:].astype(np.float32)
+        if cls_part.shape[1] == 1:
+            scores = cls_part[:, 0]
+            cls_ids = np.zeros(len(scores), dtype=np.int32)
+        else:
+            cls_ids = np.argmax(cls_part, axis=1).astype(np.int32)
+            scores = cls_part[np.arange(len(cls_part)), cls_ids]
+        cls_ids = self.cls_remap[cls_ids]
+
+        keep = self._loose_conf_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)
+
+        pad_w, pad_h = pad
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+
+        boxes, scores, cls_ids = self._filter_sane_boxes(
+            boxes, scores, cls_ids, orig_size
+        )
+        if len(boxes) == 0:
+            return []
+
+        keep = self._conf_filter_mask(scores, cls_ids)
+        boxes = boxes[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+        if len(boxes) == 0:
+            return []
+
+        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
+        return self._build_results(boxes, scores, cls_ids)
+
+    def _postprocess(
+        self,
+        output: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        if output.ndim == 2 and output.shape[1] >= 6:
+            return self._decode_final_dets(output, ratio, pad, orig_size)
+        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 6:
+            return self._decode_final_dets(output, ratio, pad, orig_size)
+        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 = (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}"
+            )
+
+        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
+        return self._postprocess(outputs[0], ratio, pad, orig_size)
+
+    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        """Horizontal-flip TTA.
+
+        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 (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(
+                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 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)
+
+        hard_keep = self._per_class_hard_nms(coords, scores, cls_ids, self.iou_thres)
+        if len(hard_keep) == 0:
+            return []
+        kept_coords = coords[hard_keep]
+        kept_scores = scores[hard_keep]
+        kept_cls = cls_ids[hard_keep]
+        kept_coords, kept_scores, kept_cls = self._nested_zone_filter(
+            kept_coords, kept_scores, kept_cls
+        )
+        if len(kept_coords) == 0:
+            return []
+        if len(kept_scores) > self.max_det:
+            top = np.argsort(-kept_scores)[: self.max_det]
+            kept_coords = kept_coords[top]
+            kept_scores = kept_scores[top]
+            kept_cls = kept_cls[top]
+
+        boosted = self._max_score_per_cluster(
+            kept_coords, kept_cls,
+            coords, scores, cls_ids, self.iou_thres,
+        )
+        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]:
+        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)
+            except Exception as 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,
+                    boxes=boxes,
+                    keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
+                )
+            )
+        return results