miner.py

import os

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"]
    # FALLBACK order the model emits classes in -- remapped to `class_names`
    # index by `self.cls_remap` (built in __init__). The authoritative order
    # is read from the ONNX `names` metadata that Ultralytics embeds at
    # export time (ships inside weights.onnx), so a retrained model with a
    # different class order is remapped correctly without code changes.
    # Used only when that metadata is missing or unparsable.
    _model_class_order = ["fire", "fire extinguisher", "smoke"]

    iou_thres = 0.55
    cross_iou_thresh = 0.8
    max_det = 150

    # Per-class confidence thresholds. Higher = fewer FP for that class.
    # Indexed by class_names order: [fire, smoke, fire_extinguisher].
    _conf_thres_array = np.array(
        [0.25, 0.3, 0.25], 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). Fire and smoke get a small bonus (variable
    # appearance); fire extinguisher does not (distinctive object, leave FP
    # control strict).
    _bonus_array = np.array(
        [0.1, 0.15, 0.1], 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

    # Same-class merge: two boxes whose intersection covers at least this
    # fraction of the SMALLER box are treated as the same object and replaced
    # by their union. Catches nested boxes (IoU below the NMS threshold) and
    # fragmented detections. Per-class because the risk differs:
    #   smoke -- diffuse plumes fragment a lot, so a moderate threshold helps.
    #   fire  -- separate flames must stay separate, so keep this HIGH (only a
    #            tight core nested inside a looser flame box merges). Set to a
    #            value > 1.0 to disable fire merging entirely.
    # Fire merge is DISABLED by default (1.01): measured on the fire-29-val1024
    # val split it cost fire AP (0.751 -> 0.742, composite 0.8888 -> 0.8874)
    # because the nested core+flame boxes it collapses were scoring as separate
    # true positives. Lower it to ~0.8 to enable, and re-measure with
    # verify_filters.py / tune_miner.py after a retrain -- a model whose fire
    # boxes fragment more (or live-SAM3 GT that draws fuller flames) could flip
    # the result.
    smoke_merge_overlap = 0.8
    fire_merge_overlap = 1.01

    # Fire containment suppression: when two FIRE boxes overlap on one object
    # (intersection >= this fraction of the SMALLER box) keep the HIGHER-conf
    # box and drop the other -- unchanged geometry, unlike the union merge
    # above. This catches the nested core+flame duplicate that per-class NMS
    # (IoU-based, iou_thres) leaves behind. Set > 1.0 to disable.
    # DISABLED by default (1.01): measured on fire-29-val1024 it cost fire AP
    # (0.751 -> 0.743, composite 0.8888 -> 0.8877). Cause: GT fire boxes almost
    # never overlap (1 pair in 416), so each nested model pair has one TP + one
    # FP, but the higher-CONF box isn't always the one matching GT at IoU 0.5 --
    # so keeping it can drop the real match, and score-ordered AP already
    # tolerates the duplicate. Lower to ~0.8 to enable; re-measure after a
    # retrain or against live-SAM3 GT, which may differ.
    fire_suppress_overlap = 0.88

    # ── Low-confidence color-prior FP filters ───────────────────────────────
    # Ported from the firedetect1007 miner's color checks, but applied ONLY to
    # the borderline confidence band (just above each per-class threshold) and
    # ONLY on color frames. A fire/extinguisher detection there is dropped when
    # its pixels clearly do not match the expected appearance: warm/bright for
    # fire, red for extinguisher. High-confidence detections are never touched.
    #
    # The reference miner ran these unconditionally -- a BUG on this validator,
    # which feeds some frames as grayscale (a true red extinguisher is gray
    # there, so a red test would wrongly delete it). We skip the filter when the
    # ROI is near-grayscale, so it never fires on those frames.
    #
    # Tunable: set a max-conf gate to 0.0 to disable that filter. After a model
    # retrain, re-validate these with tune_miner.py (the gates are relative to
    # the per-class thresholds, so they move when those move).
    fire_color_filter_max_conf = 0.45      # only fire boxes in (thresh, 0.45]
    fire_ext_color_filter_max_conf = 0.40  # only ext boxes in (thresh, 0.40]
    color_filter_min_saturation = 0.06     # skip filter if ROI is near-grayscale

    # ── Corroboration FP filters (optional; OFF by default) ─────────────────
    # Ported in spirit from firedetect1007. Both REMOVE borderline boxes that
    # lack support -- a precision play for the validator's FP pillar. OFF by
    # default because, unlike the color priors, they can also drop true
    # positives; enable + sweep with verify_filters.py and keep only the
    # settings that raise the measured composite. A max-conf gate of 0.0
    # disables the corresponding filter.
    #   edge filter: drop boxes touching the frame border in a low-conf band
    #     (the validator scales/crops, so border-hugging boxes are often the
    #     truncated remains of an object whose body is off-frame).
    #   tta view filter: drop low-conf boxes that appear in only ONE of the two
    #     horizontal-flip TTA views (a real object is usually seen in both).
    use_edge_filter = False
    edge_filter_max_conf = 0.0      # drop edge-touching boxes with conf <= this
    edge_tol = 2.0                  # px from the border counted as "on edge"
    use_tta_view_filter = False
    tta_view_filter_max_conf = 0.0  # drop single-view boxes with conf <= this
    tta_view_iou_thresh = 0.5       # IoU for "same object seen in both views"

    def __init__(self, path_hf_repo: Path) -> None:
        model_path = path_hf_repo / "weights.onnx"
        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
        sess_options.intra_op_num_threads = 2
        sess_options.inter_op_num_threads = 1

        try:
            self.session = ort.InferenceSession(
                str(model_path),
                sess_options=sess_options,
                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
            )
            print("✅ Created ORT session with preferred CUDA provider list")
        except Exception as e:
            print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}")
            self.session = ort.InferenceSession(
                str(model_path),
                sess_options=sess_options,
                providers=["CPUExecutionProvider"],
            )

        print("ORT session providers:", self.session.get_providers())

        # Build cls_remap: for each model-emit index i,
        #   cls_remap[i] = self.class_names.index(model_class_order[i])
        # i.e. converts a model-side class id into the canonical class id
        # that downstream code (BoundingBox.cls_id, validator) 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)
            print(f"cls order: no usable ONNX metadata, FALLBACK {model_class_order}")
        else:
            print(f"cls order: from ONNX metadata {model_class_order}")
        self.cls_remap = np.array(
            [self.class_names.index(n) for n in model_class_order],
            dtype=np.int32,
        )

        for inp in self.session.get_inputs():
            print("INPUT:", inp.name, inp.shape, inp.type)
        for out in self.session.get_outputs():
            print("OUTPUT:", out.name, out.shape, out.type)

        self.input_name = self.session.get_inputs()[0].name
        self.output_names = [output.name for output in self.session.get_outputs()]
        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 = False

        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()
            )
        ))

        self._warmup()

    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")
        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

    def _read_model_class_order(self) -> list[str] | None:
        """Read the model's class order from Ultralytics ONNX metadata.

        Returns the class names ordered by model-emit index, or None when
        metadata is missing/unparsable or doesn't match `class_names` as a
        set (in which case the static _model_class_order fallback is used).
        """
        try:
            import ast

            meta = self.session.get_modelmeta().custom_metadata_map
            names = ast.literal_eval(meta["names"])  # e.g. {0: 'fire', ...}
            if isinstance(names, dict):
                order = [str(names[i]) for i in sorted(names)]
            else:
                order = [str(n) for n in names]
        except Exception as e:
            print(f"cls order: could not read ONNX names metadata ({e})")
            return None
        if sorted(order) != sorted(self.class_names):
            print(
                f"cls order: ONNX names {order} do not match expected classes "
                f"{self.class_names}; ignoring metadata"
            )
            return None
        return order

    def _letterbox(
        self,
        image: ndarray,
        new_shape: tuple[int, int],
        color=(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)
        )
        # Fused scale(1/255) + BGR->RGB swap + HWC->NCHW + contiguous float32 in
        # one optimized OpenCV call. Bit-identical (max abs diff 6e-8) to the
        # prior cvtColor + astype/255 + transpose + ascontiguousarray chain, but
        # ~half the preprocess time (preprocess is ~12% of predict_batch).
        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:
        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)

    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).
        """
        n = len(boxes)
        if n <= 1:
            return boxes, scores, cls_ids
        boxes = np.asarray(boxes, dtype=np.float32)
        scores = np.asarray(scores, dtype=np.float32)
        cls_ids = np.asarray(cls_ids, dtype=np.int32)
        areas = (np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) *
                 np.maximum(0.0, boxes[:, 3] - boxes[:, 1]))
        margins = scores - self._conf_thres_array[cls_ids]
        order = np.lexsort((-areas, -margins))
        suppressed = np.zeros(n, dtype=bool)
        keep: 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
            suppressed |= dup
        keep_idx = np.array(keep, dtype=np.intp)
        return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx]

    def _merge_class_boxes(
        self,
        boxes: np.ndarray,
        scores: np.ndarray,
        cls_ids: np.ndarray,
        target_cls: int,
        overlap: float,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        """Merge overlapping detections of ONE class into single boxes.

        Two same-class boxes whose intersection covers >= `overlap` of the
        SMALLER box are treated as one object and replaced by their union with
        the max confidence of the pair. Repeats until no pair merges, so chains
        of fragments collapse. `overlap` is intersection-over-minimum-area, so
        only nested / heavily-overlapping boxes merge -- two spatially separate
        objects (low mutual overlap) are never fused. `overlap > 1.0` disables.
        """
        if overlap > 1.0:
            return boxes, scores, cls_ids
        idx = np.where(cls_ids == target_cls)[0]
        if len(idx) <= 1:
            return boxes, scores, cls_ids

        sb = boxes[idx].astype(np.float32).tolist()
        ss = scores[idx].astype(np.float32).tolist()
        merged_any = True
        while merged_any and len(sb) > 1:
            merged_any = False
            for i in range(len(sb)):
                for j in range(i + 1, len(sb)):
                    a, b = sb[i], sb[j]
                    ix1 = max(a[0], b[0])
                    iy1 = max(a[1], b[1])
                    ix2 = min(a[2], b[2])
                    iy2 = min(a[3], b[3])
                    inter = max(0.0, ix2 - ix1) * max(0.0, iy2 - iy1)
                    area_a = max(0.0, a[2] - a[0]) * max(0.0, a[3] - a[1])
                    area_b = max(0.0, b[2] - b[0]) * max(0.0, b[3] - b[1])
                    smaller = min(area_a, area_b)
                    if inter / (smaller + 1e-7) >= overlap:
                        sb[i] = [
                            min(a[0], b[0]), min(a[1], b[1]),
                            max(a[2], b[2]), max(a[3], b[3]),
                        ]
                        ss[i] = max(ss[i], ss[j])
                        del sb[j]
                        del ss[j]
                        merged_any = True
                        break
                if merged_any:
                    break

        other = cls_ids != target_cls
        new_boxes = np.concatenate(
            [boxes[other].astype(np.float32),
             np.array(sb, dtype=np.float32).reshape(-1, 4)]
        )
        new_scores = np.concatenate(
            [scores[other].astype(np.float32),
             np.array(ss, dtype=np.float32)]
        )
        new_cls = np.concatenate(
            [cls_ids[other].astype(np.int32),
             np.full(len(sb), target_cls, dtype=np.int32)]
        )
        return new_boxes, new_scores, new_cls

    def _suppress_contained_lower_conf(
        self,
        boxes: np.ndarray,
        scores: np.ndarray,
        cls_ids: np.ndarray,
        target_cls: int,
        overlap: float,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        """For one class, when two boxes overlap (intersection >= `overlap` of
        the smaller box) keep the higher-confidence box and drop the other.
        Geometry is never changed -- only the redundant lower-conf box is
        removed. `overlap > 1.0` disables."""
        if overlap > 1.0:
            return boxes, scores, cls_ids
        idx = np.where(cls_ids == target_cls)[0]
        if len(idx) <= 1:
            return boxes, scores, cls_ids

        order = idx[np.argsort(-scores[idx])]  # highest confidence first
        remove: set[int] = set()
        for a in range(len(order)):
            i = int(order[a])
            if i in remove:
                continue
            bi = boxes[i]
            area_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
            for b in range(a + 1, len(order)):
                j = int(order[b])
                if j in remove:
                    continue
                bj = boxes[j]
                ix1 = max(bi[0], bj[0]); iy1 = max(bi[1], bj[1])
                ix2 = min(bi[2], bj[2]); iy2 = min(bi[3], bj[3])
                inter = max(0.0, ix2 - ix1) * max(0.0, iy2 - iy1)
                if inter <= 0.0:
                    continue
                area_j = max(1e-7, float((bj[2] - bj[0]) * (bj[3] - bj[1])))
                if inter / (min(area_i, area_j) + 1e-7) >= overlap:
                    remove.add(j)  # j is the lower-confidence box (order desc)
        if not remove:
            return boxes, scores, cls_ids
        keep = np.array(
            [k not in remove for k in range(len(boxes))], dtype=bool
        )
        return boxes[keep], scores[keep], cls_ids[keep]

    def _merge_same_class_boxes(
        self,
        boxes: np.ndarray,
        scores: np.ndarray,
        cls_ids: np.ndarray,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        """Resolve nested / fragmented same-object detections, per class.

        Smoke: diffuse plumes fragment into nested boxes NMS can't collapse, so
        they are UNION-merged (smoke_merge_overlap).
        Fire: a tight hot-core box and a looser flame box are the same flame;
        keep the HIGHER-confidence one and drop the other (fire_suppress_overlap),
        which leaves geometry intact. The union-merge variant (fire_merge_overlap)
        is also available but measured worse, so it is disabled by default.
        """
        boxes, scores, cls_ids = self._merge_class_boxes(
            boxes, scores, cls_ids,
            self.class_names.index("smoke"), self.smoke_merge_overlap,
        )
        boxes, scores, cls_ids = self._merge_class_boxes(
            boxes, scores, cls_ids,
            self.class_names.index("fire"), self.fire_merge_overlap,
        )
        boxes, scores, cls_ids = self._suppress_contained_lower_conf(
            boxes, scores, cls_ids,
            self.class_names.index("fire"), self.fire_suppress_overlap,
        )
        return boxes, scores, cls_ids

    # Back-compat alias (older callers / tune_miner referenced this name).
    def _merge_smoke_boxes(
        self,
        boxes: np.ndarray,
        scores: np.ndarray,
        cls_ids: np.ndarray,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        return self._merge_same_class_boxes(boxes, scores, cls_ids)

    @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 _conf_filter_mask(
        self, scores: np.ndarray, cls_ids: np.ndarray
    ) -> np.ndarray:
        """Boolean keep-mask: score >= per-class threshold, with a per-class
        rescue -- if a class has zero boxes passing, admit its top-1 candidate
        when its score >= (per-class threshold - per-class bonus)."""
        if len(scores) == 0:
            return np.zeros(0, dtype=bool)
        thr = self._conf_thres_array[cls_ids]
        keep = scores >= thr
        for c in np.unique(cls_ids):
            b = float(self._bonus_array[c])
            if b <= 0.0:
                continue
            cm = cls_ids == c
            if keep[cm].any():
                continue
            idx = np.where(cm)[0]
            top = int(idx[int(np.argmax(scores[idx]))])
            if scores[top] >= self._conf_thres_array[c] - b:
                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 -> 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]
        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
            )
        if len(boxes) > 1:
            boxes, scores, cls_ids = self._merge_same_class_boxes(boxes, scores, cls_ids)
        return boxes, scores, cls_ids

    @staticmethod
    def _roi_for_box(image: np.ndarray, box: BoundingBox) -> np.ndarray | None:
        """Clip a BoundingBox to the image and return its BGR pixel ROI."""
        h, w = image.shape[:2]
        x1 = max(0, int(math.floor(box.x1)))
        y1 = max(0, int(math.floor(box.y1)))
        x2 = min(w, int(math.ceil(box.x2)))
        y2 = min(h, int(math.ceil(box.y2)))
        if x2 <= x1 or y2 <= y1:
            return None
        roi = image[y1:y2, x1:x2]
        return roi if roi.size else None

    def _roi_is_near_grayscale(self, roi: np.ndarray) -> bool:
        """True if the ROI carries almost no color (validator grayscale frame).
        On such ROIs the color priors are skipped so they can't delete valid
        red/warm objects that have been stripped of color."""
        mx = roi.max(axis=2).astype(np.float32)
        mn = roi.min(axis=2).astype(np.float32)
        sat = (mx - mn) / (mx + 1e-6)
        return float(sat.mean()) < self.color_filter_min_saturation

    @staticmethod
    def _passes_fire_color(roi: np.ndarray) -> bool:
        """Fire is warm and/or has a bright hotspot. ROI is BGR."""
        blue = roi[:, :, 0].astype(np.float32)
        green = roi[:, :, 1].astype(np.float32)
        red = roi[:, :, 2].astype(np.float32)
        mean_r = float(np.mean(red))
        max_rgb = float(max(np.max(red), np.max(green), np.max(blue)))
        bright_frac = float(np.mean(np.max(roi, axis=2) >= 150))
        # A bright hotspot is fire-like even with little hue (also covers the
        # near-white core of an intense flame).
        if max_rgb >= 200.0 and bright_frac >= 0.01:
            return True
        warm = (red > green + 10.0) & (red > blue + 10.0)
        warm_frac = float(np.mean(warm))
        r_minus_g = mean_r - float(np.mean(green))
        if warm_frac >= 0.05 and (
            max_rgb >= 120.0 or mean_r >= 120.0 or warm_frac >= 0.15
        ):
            return True
        if bright_frac >= 0.12 and r_minus_g >= 2.0:
            return True
        return False

    @staticmethod
    def _passes_fire_ext_red_color(roi: np.ndarray) -> bool:
        """Fire extinguishers are red. ROI is BGR. Lenient: only clearly
        cool/green/blue or very dark regions fail."""
        blue = roi[:, :, 0].astype(np.float32)
        green = roi[:, :, 1].astype(np.float32)
        red = roi[:, :, 2].astype(np.float32)
        red_dom = float(np.mean((red > green + 10.0) & (red > blue + 10.0)))
        if red_dom >= 0.03:
            return True
        if (float(np.mean(red)) - float(np.mean(green))) >= 0.0 and \
                float(np.mean(red)) >= 50.0:
            return True
        return False

    def _remove_edge_low_conf(
        self, results: list[BoundingBox], orig_size: tuple[int, int]
    ) -> list[BoundingBox]:
        """Drop border-hugging boxes in the low-confidence band."""
        if (
            not self.use_edge_filter
            or self.edge_filter_max_conf <= 0.0
            or not results
        ):
            return results
        w, h = orig_size
        tol = self.edge_tol
        out: list[BoundingBox] = []
        for b in results:
            on_edge = (
                b.x1 <= tol
                or b.y1 <= tol
                or b.x2 >= w - 1 - tol
                or b.y2 >= h - 1 - tol
            )
            if on_edge and b.conf <= self.edge_filter_max_conf:
                continue
            out.append(b)
        return out

    def _views_corroborated(
        self,
        post_boxes: np.ndarray,
        post_cls: np.ndarray,
        full_boxes: np.ndarray,
        full_cls: np.ndarray,
        full_views: np.ndarray,
        iou_thresh: float,
    ) -> np.ndarray:
        """For each post-NMS box, True if same-class detections from >= 2
        distinct TTA views overlap it (IoU >= iou_thresh) in the full union."""
        n = len(post_boxes)
        if n == 0:
            return np.zeros(0, dtype=bool)
        full_areas = (np.maximum(0.0, full_boxes[:, 2] - full_boxes[:, 0]) *
                      np.maximum(0.0, full_boxes[:, 3] - full_boxes[:, 1]))
        out = np.zeros(n, dtype=bool)
        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)
            mask = (iou >= iou_thresh) & (full_cls == post_cls[i])
            if np.any(mask):
                out[i] = len(np.unique(full_views[mask])) >= 2
        return out

    def _filter_low_conf_by_color(
        self, image: np.ndarray, results: list[BoundingBox]
    ) -> list[BoundingBox]:
        """Drop borderline fire / extinguisher detections whose pixels clearly
        contradict the class's expected color. No-op on near-grayscale ROIs and
        on detections above the per-class color-filter conf gate."""
        if not results:
            return results
        cls_fire = self.class_names.index("fire")
        cls_ext = self.class_names.index("fire extinguisher")
        out: list[BoundingBox] = []
        for box in results:
            check_fire = (
                box.cls_id == cls_fire
                and box.conf <= self.fire_color_filter_max_conf
            )
            check_ext = (
                box.cls_id == cls_ext
                and box.conf <= self.fire_ext_color_filter_max_conf
            )
            if not check_fire and not check_ext:
                out.append(box)
                continue
            roi = self._roi_for_box(image, box)
            if roi is None or self._roi_is_near_grayscale(roi):
                out.append(box)
                continue
            if check_fire and not self._passes_fire_color(roi):
                continue
            if check_ext and not self._passes_fire_ext_red_color(roi):
                continue
            out.append(box)
        return out

    @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._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
        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 []

        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._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)

        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 []

        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.
          6. Smoke merge: overlapping / nested smoke boxes collapse into
             their union (one box per smoke object).
        """
        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)
        # view_id 0 = original, 1 = horizontal flip (mapped back to orig coords)
        view_ids = np.array(
            [0] * len(boxes_orig) + [1] * len(boxes_flip), dtype=np.int32
        )

        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]

        # Optional: drop low-conf detections seen in only one TTA view.
        if (
            self.use_tta_view_filter
            and self.tta_view_filter_max_conf > 0.0
            and len(kept_coords) > 0
        ):
            corrob = self._views_corroborated(
                kept_coords, kept_cls, coords, cls_ids, view_ids,
                self.tta_view_iou_thresh,
            )
            keep = ~((boosted <= self.tta_view_filter_max_conf) & (~corrob))
            kept_coords = kept_coords[keep]
            boosted = boosted[keep]
            kept_cls = kept_cls[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
            )
        if len(kept_coords) > 1:
            kept_coords, boosted, kept_cls = self._merge_same_class_boxes(
                kept_coords, boosted, kept_cls
            )

        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)
                # Color-prior + edge FP filters on the merged result, in
                # original-image coords. Single insertion point so they run once
                # per frame for both the TTA and non-TTA paths.
                if isinstance(image, np.ndarray) and image.ndim == 3:
                    boxes = self._filter_low_conf_by_color(image, boxes)
                    boxes = self._remove_edge_low_conf(
                        boxes, (image.shape[1], image.shape[0])
                    )
            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