innomium/fire-detection

1from __future__ import annotations2 3import math4from pathlib import Path5 6import cv27import numpy as np8import onnxruntime as ort9from numpy import ndarray10from pydantic import BaseModel11 12MODEL_DIR = Path(__file__).resolve().parent13 14class BoundingBox(BaseModel):15    x1: int16    y1: int17    x2: int18    y2: int19    cls_id: int20    conf: float21 22 23class TVFrameResult(BaseModel):24    frame_id: int25    boxes: list[BoundingBox]26    keypoints: list[tuple[int, int]]27 28 29class FireDetector:30    """ONNX Runtime miner for fire / smoke / fire_extinguisher detection.31 32    Strategy (ported from offense miner):33      - per-class confidence threshold with per-class rescue bonus34      - per-class hard NMS, then cross-class dedup35      - horizontal-flip TTA with full-set cluster score boost36    Plus fire001 specifics: class remap, sanity-box filter, TTA toggle.37    """38 39    class_names = ["fire", "smoke", "fire extinguisher"]40    # FALLBACK order the model emits classes in -- remapped to `class_names`41    # index by `self.cls_remap` (built in __init__). The authoritative order42    # is read from the ONNX `names` metadata that Ultralytics embeds at43    # export time (ships inside weights.onnx), so a retrained model with a44    # different class order is remapped correctly without code changes.45    # Used only when that metadata is missing or unparsable.46    _model_class_order = ["fire", "fire extinguisher", "smoke"]47 48    iou_thres = 0.5549    cross_iou_thresh = 0.850    max_det = 15051 52    # Per-class confidence thresholds. Higher = fewer FP for that class.53    # Indexed by class_names order: [fire, smoke, fire_extinguisher].54    _conf_thres_array = np.array(55        [0.15, 0.30, 0.23], dtype=np.float3256    )57    # Per-class rescue bonus. If a class has ZERO boxes passing the threshold58    # in a frame, its top-1 candidate is admitted when its score is at least59    # (threshold - bonus). Fire and smoke get a small bonus (variable60    # appearance); fire extinguisher does not (distinctive object, leave FP61    # control strict).62    _bonus_array = np.array(63        [0.02, 0.1, 0.1], dtype=np.float3264    )65 66    # Box sanity filter (fire001-specific FP reduction): drop tiny / degenerate67    # / image-spanning / extreme aspect ratio boxes.68    min_box_area = 14 * 1469    min_side = 870    max_aspect_ratio = 8.071 72    # Same-class merge: two boxes whose intersection covers at least this73    # fraction of the SMALLER box are treated as the same object and replaced74    # by their union. Catches nested boxes (IoU below the NMS threshold) and75    # fragmented detections. Per-class because the risk differs:76    #   smoke -- diffuse plumes fragment a lot, so a moderate threshold helps.77    #   fire  -- separate flames must stay separate, so keep this HIGH (only a78    #            tight core nested inside a looser flame box merges). Set to a79    #            value > 1.0 to disable fire merging entirely.80    # Fire merge is DISABLED by default (1.01): measured on the fire-29-val102481    # val split it cost fire AP (0.751 -> 0.742, composite 0.8888 -> 0.8874)82    # because the nested core+flame boxes it collapses were scoring as separate83    # true positives. Lower it to ~0.8 to enable, and re-measure with84    # verify_filters.py / tune_miner.py after a retrain -- a model whose fire85    # boxes fragment more (or live-SAM3 GT that draws fuller flames) could flip86    # the result.87    smoke_merge_overlap = 0.888    fire_merge_overlap = 1.0189 90    # Fire containment suppression: when two FIRE boxes overlap on one object91    # (intersection >= this fraction of the SMALLER box) keep the HIGHER-conf92    # box and drop the other -- unchanged geometry, unlike the union merge93    # above. This catches the nested core+flame duplicate that per-class NMS94    # (IoU-based, iou_thres) leaves behind. Set > 1.0 to disable.95    # DISABLED by default (1.01): measured on fire-29-val1024 it cost fire AP96    # (0.751 -> 0.743, composite 0.8888 -> 0.8877). Cause: GT fire boxes almost97    # never overlap (1 pair in 416), so each nested model pair has one TP + one98    # FP, but the higher-CONF box isn't always the one matching GT at IoU 0.5 --99    # so keeping it can drop the real match, and score-ordered AP already100    # tolerates the duplicate. Lower to ~0.8 to enable; re-measure after a101    # retrain or against live-SAM3 GT, which may differ.102    fire_suppress_overlap = 0.88103 104    # ── Low-confidence color-prior FP filters ───────────────────────────────105    # Ported from the firedetect1007 miner's color checks, but applied ONLY to106    # the borderline confidence band (just above each per-class threshold) and107    # ONLY on color frames. A fire/extinguisher detection there is dropped when108    # its pixels clearly do not match the expected appearance: warm/bright for109    # fire, red for extinguisher. High-confidence detections are never touched.110    #111    # The reference miner ran these unconditionally -- a BUG on this validator,112    # which feeds some frames as grayscale (a true red extinguisher is gray113    # there, so a red test would wrongly delete it). We skip the filter when the114    # ROI is near-grayscale, so it never fires on those frames.115    #116    # Tunable: set a max-conf gate to 0.0 to disable that filter. After a model117    # retrain, re-validate these with tune_miner.py (the gates are relative to118    # the per-class thresholds, so they move when those move).119    fire_color_filter_max_conf = 0.45      # only fire boxes in (thresh, 0.45]120    fire_ext_color_filter_max_conf = 0.40  # only ext boxes in (thresh, 0.40]121    color_filter_min_saturation = 0.06     # skip filter if ROI is near-grayscale122 123    # ── Corroboration FP filters (optional; OFF by default) ─────────────────124    # Ported in spirit from firedetect1007. Both REMOVE borderline boxes that125    # lack support -- a precision play for the validator's FP pillar. OFF by126    # default because, unlike the color priors, they can also drop true127    # positives; enable + sweep with verify_filters.py and keep only the128    # settings that raise the measured composite. A max-conf gate of 0.0129    # disables the corresponding filter.130    #   edge filter: drop boxes touching the frame border in a low-conf band131    #     (the validator scales/crops, so border-hugging boxes are often the132    #     truncated remains of an object whose body is off-frame).133    #   tta view filter: drop low-conf boxes that appear in only ONE of the two134    #     horizontal-flip TTA views (a real object is usually seen in both).135    use_edge_filter = False136    edge_filter_max_conf = 0.0      # drop edge-touching boxes with conf <= this137    edge_tol = 2.0                  # px from the border counted as "on edge"138    use_tta_view_filter = False139    tta_view_filter_max_conf = 0.0  # drop single-view boxes with conf <= this140    tta_view_iou_thresh = 0.5       # IoU for "same object seen in both views"141 142    def __init__(self, model_dir: Path | str | None = None) -> None:143        model_dir = Path(model_dir) if model_dir is not None else MODEL_DIR144        model_path = model_dir / "weights.onnx"145        print("ORT version:", ort.__version__)146 147        try:148            ort.preload_dlls()149            print("✅ onnxruntime.preload_dlls() success")150        except Exception as e:151            print(f"⚠️ preload_dlls failed: {e}")152 153        print("ORT available providers BEFORE session:", ort.get_available_providers())154 155        sess_options = ort.SessionOptions()156        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL157        sess_options.intra_op_num_threads = 2158        sess_options.inter_op_num_threads = 1159        sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL160 161        try:162            self.session = ort.InferenceSession(163                str(model_path),164                sess_options=sess_options,165                providers=["CPUExecutionProvider"],166            )167        except Exception as e:168            self.session = ort.InferenceSession(169                str(model_path),170                sess_options=sess_options,171                providers=["CPUExecutionProvider"],172            )173 174        print("ORT session providers:", self.session.get_providers())175 176        # Build cls_remap: for each model-emit index i,177        #   cls_remap[i] = self.class_names.index(model_class_order[i])178        # i.e. converts a model-side class id into the canonical class id179        # that downstream code (BoundingBox.cls_id, validator) expects.180        # The model-side order comes from the ONNX metadata when available,181        # else falls back to the static _model_class_order.182        model_class_order = self._read_model_class_order()183        if model_class_order is None:184            model_class_order = list(self._model_class_order)185            print(f"cls order: no usable ONNX metadata, FALLBACK {model_class_order}")186        else:187            print(f"cls order: from ONNX metadata {model_class_order}")188        self.cls_remap = np.array(189            [self.class_names.index(n) for n in model_class_order],190            dtype=np.int32,191        )192 193        for inp in self.session.get_inputs():194            print("INPUT:", inp.name, inp.shape, inp.type)195        for out in self.session.get_outputs():196            print("OUTPUT:", out.name, out.shape, out.type)197 198        self.input_name = self.session.get_inputs()[0].name199        self.output_names = [output.name for output in self.session.get_outputs()]200        self.input_shape = self.session.get_inputs()[0].shape201 202        self.input_height = self._safe_dim(self.input_shape[2], default=1280)203        self.input_width = self._safe_dim(self.input_shape[3], default=1280)204 205        self.use_tta = False206 207        print(f"✅ ONNX model loaded from: {model_path}")208        print(f"✅ ONNX providers: {self.session.get_providers()}")209        print(f"✅ ONNX input: name={self.input_name}, shape={self.input_shape}")210        print("per-class conf: " + ", ".join(211            f"{n}={t:.3f}" for n, t in zip(212                self.class_names, self._conf_thres_array.tolist()213            )214        ))215 216        self._warmup()217 218    def _warmup(self, iters: int = 3) -> None:219        try:220            dummy = np.zeros((720, 1280, 3), dtype=np.uint8)221            for _ in range(max(1, iters)):222                self.predict_batch(batch_images=[dummy], offset=0, n_keypoints=0)223            print(f"✅ warmup: {iters} dummy predict_batch call(s) done")224        except Exception as e:225            print(f"⚠️ warmup skipped: {e}")226 227    def __repr__(self) -> str:228        return (229            f"ONNXRuntime(session={type(self.session).__name__}, "230            f"providers={self.session.get_providers()})"231        )232 233    @staticmethod234    def _safe_dim(value, default: int) -> int:235        return value if isinstance(value, int) and value > 0 else default236 237    def _read_model_class_order(self) -> list[str] | None:238        """Read the model's class order from Ultralytics ONNX metadata.239 240        Returns the class names ordered by model-emit index, or None when241        metadata is missing/unparsable or doesn't match `class_names` as a242        set (in which case the static _model_class_order fallback is used).243        """244        try:245            import ast246 247            meta = self.session.get_modelmeta().custom_metadata_map248            names = ast.literal_eval(meta["names"])  # e.g. {0: 'fire', ...}249            if isinstance(names, dict):250                order = [str(names[i]) for i in sorted(names)]251            else:252                order = [str(n) for n in names]253        except Exception as e:254            print(f"cls order: could not read ONNX names metadata ({e})")255            return None256        if sorted(order) != sorted(self.class_names):257            print(258                f"cls order: ONNX names {order} do not match expected classes "259                f"{self.class_names}; ignoring metadata"260            )261            return None262        return order263 264    def _letterbox(265        self,266        image: ndarray,267        new_shape: tuple[int, int],268        color=(114, 114, 114),269    ) -> tuple[ndarray, float, tuple[float, float]]:270        h, w = image.shape[:2]271        new_w, new_h = new_shape272 273        ratio = min(new_w / w, new_h / h)274        resized_w = int(round(w * ratio))275        resized_h = int(round(h * ratio))276 277        if (resized_w, resized_h) != (w, h):278            interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR279            image = cv2.resize(image, (resized_w, resized_h), interpolation=interp)280 281        dw = (new_w - resized_w) / 2.0282        dh = (new_h - resized_h) / 2.0283 284        left = int(round(dw - 0.1))285        right = int(round(dw + 0.1))286        top = int(round(dh - 0.1))287        bottom = int(round(dh + 0.1))288 289        padded = cv2.copyMakeBorder(290            image, top, bottom, left, right,291            borderType=cv2.BORDER_CONSTANT, value=color,292        )293        return padded, ratio, (dw, dh)294 295    def _preprocess(296        self, image: ndarray297    ) -> tuple[np.ndarray, float, tuple[float, float], tuple[int, int]]:298        orig_h, orig_w = image.shape[:2]299        img, ratio, pad = self._letterbox(300            image, (self.input_width, self.input_height)301        )302        # Fused scale(1/255) + BGR->RGB swap + HWC->NCHW + contiguous float32 in303        # one optimized OpenCV call. Bit-identical (max abs diff 6e-8) to the304        # prior cvtColor + astype/255 + transpose + ascontiguousarray chain, but305        # ~half the preprocess time (preprocess is ~12% of predict_batch).306        blob = cv2.dnn.blobFromImage(img, scalefactor=1.0 / 255.0, swapRB=True)307        return blob, ratio, pad, (orig_w, orig_h)308 309    @staticmethod310    def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:311        w, h = image_size312        boxes[:, 0] = np.clip(boxes[:, 0], 0, w - 1)313        boxes[:, 1] = np.clip(boxes[:, 1], 0, h - 1)314        boxes[:, 2] = np.clip(boxes[:, 2], 0, w - 1)315        boxes[:, 3] = np.clip(boxes[:, 3], 0, h - 1)316        return boxes317 318    @staticmethod319    def _xywh_to_xyxy(boxes: np.ndarray) -> np.ndarray:320        out = np.empty_like(boxes)321        out[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.0322        out[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.0323        out[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.0324        out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0325        return out326 327    @staticmethod328    def _hard_nms(329        boxes: np.ndarray, scores: np.ndarray, iou_thresh: float330    ) -> np.ndarray:331        n = len(boxes)332        if n == 0:333            return np.array([], dtype=np.intp)334        order = np.argsort(-scores)335        keep: list[int] = []336        while len(order) > 0:337            i = int(order[0])338            keep.append(i)339            if len(order) == 1:340                break341            rest = order[1:]342            xx1 = np.maximum(boxes[i, 0], boxes[rest, 0])343            yy1 = np.maximum(boxes[i, 1], boxes[rest, 1])344            xx2 = np.minimum(boxes[i, 2], boxes[rest, 2])345            yy2 = np.minimum(boxes[i, 3], boxes[rest, 3])346            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)347            a_i = (max(0.0, boxes[i, 2] - boxes[i, 0]) *348                   max(0.0, boxes[i, 3] - boxes[i, 1]))349            a_r = (np.maximum(0.0, boxes[rest, 2] - boxes[rest, 0]) *350                   np.maximum(0.0, boxes[rest, 3] - boxes[rest, 1]))351            iou = inter / (a_i + a_r - inter + 1e-7)352            order = rest[iou <= iou_thresh]353        return np.array(keep, dtype=np.intp)354 355    def _per_class_hard_nms(356        self,357        boxes: np.ndarray,358        scores: np.ndarray,359        cls_ids: np.ndarray,360        iou_thresh: float,361    ) -> np.ndarray:362        if len(boxes) == 0:363            return np.array([], dtype=np.intp)364        all_keep: list[int] = []365        for c in np.unique(cls_ids):366            mask = cls_ids == c367            indices = np.where(mask)[0]368            keep = self._hard_nms(boxes[mask], scores[mask], iou_thresh)369            all_keep.extend(indices[keep].tolist())370        all_keep.sort()371        return np.array(all_keep, dtype=np.intp)372 373    def _cross_class_dedup_op(374        self,375        boxes: np.ndarray,376        scores: np.ndarray,377        cls_ids: np.ndarray,378        iou_thresh: float,379    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:380        """Remove near-duplicate boxes across classes.381 382        Order candidates by (score - per_class_threshold) margin, then by area;383        keep the highest, suppress every other box with IoU > iou_thresh.384        This suppresses the case where the same physical object is detected385        as multiple classes (e.g. fire vs smoke on the same flames).386        """387        n = len(boxes)388        if n <= 1:389            return boxes, scores, cls_ids390        boxes = np.asarray(boxes, dtype=np.float32)391        scores = np.asarray(scores, dtype=np.float32)392        cls_ids = np.asarray(cls_ids, dtype=np.int32)393        areas = (np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) *394                 np.maximum(0.0, boxes[:, 3] - boxes[:, 1]))395        margins = scores - self._conf_thres_array[cls_ids]396        order = np.lexsort((-areas, -margins))397        suppressed = np.zeros(n, dtype=bool)398        keep: list[int] = []399        for i in order:400            if suppressed[i]:401                continue402            keep.append(int(i))403            bi = boxes[i]404            xx1 = np.maximum(bi[0], boxes[:, 0])405            yy1 = np.maximum(bi[1], boxes[:, 1])406            xx2 = np.minimum(bi[2], boxes[:, 2])407            yy2 = np.minimum(bi[3], boxes[:, 3])408            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)409            a_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))410            iou = inter / (a_i + areas - inter + 1e-7)411            dup = iou > iou_thresh412            dup[i] = False413            suppressed |= dup414        keep_idx = np.array(keep, dtype=np.intp)415        return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx]416 417    def _merge_class_boxes(418        self,419        boxes: np.ndarray,420        scores: np.ndarray,421        cls_ids: np.ndarray,422        target_cls: int,423        overlap: float,424    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:425        """Merge overlapping detections of ONE class into single boxes.426 427        Two same-class boxes whose intersection covers >= `overlap` of the428        SMALLER box are treated as one object and replaced by their union with429        the max confidence of the pair. Repeats until no pair merges, so chains430        of fragments collapse. `overlap` is intersection-over-minimum-area, so431        only nested / heavily-overlapping boxes merge -- two spatially separate432        objects (low mutual overlap) are never fused. `overlap > 1.0` disables.433        """434        if overlap > 1.0:435            return boxes, scores, cls_ids436        idx = np.where(cls_ids == target_cls)[0]437        if len(idx) <= 1:438            return boxes, scores, cls_ids439 440        sb = boxes[idx].astype(np.float32).tolist()441        ss = scores[idx].astype(np.float32).tolist()442        merged_any = True443        while merged_any and len(sb) > 1:444            merged_any = False445            for i in range(len(sb)):446                for j in range(i + 1, len(sb)):447                    a, b = sb[i], sb[j]448                    ix1 = max(a[0], b[0])449                    iy1 = max(a[1], b[1])450                    ix2 = min(a[2], b[2])451                    iy2 = min(a[3], b[3])452                    inter = max(0.0, ix2 - ix1) * max(0.0, iy2 - iy1)453                    area_a = max(0.0, a[2] - a[0]) * max(0.0, a[3] - a[1])454                    area_b = max(0.0, b[2] - b[0]) * max(0.0, b[3] - b[1])455                    smaller = min(area_a, area_b)456                    if inter / (smaller + 1e-7) >= overlap:457                        sb[i] = [458                            min(a[0], b[0]), min(a[1], b[1]),459                            max(a[2], b[2]), max(a[3], b[3]),460                        ]461                        ss[i] = max(ss[i], ss[j])462                        del sb[j]463                        del ss[j]464                        merged_any = True465                        break466                if merged_any:467                    break468 469        other = cls_ids != target_cls470        new_boxes = np.concatenate(471            [boxes[other].astype(np.float32),472             np.array(sb, dtype=np.float32).reshape(-1, 4)]473        )474        new_scores = np.concatenate(475            [scores[other].astype(np.float32),476             np.array(ss, dtype=np.float32)]477        )478        new_cls = np.concatenate(479            [cls_ids[other].astype(np.int32),480             np.full(len(sb), target_cls, dtype=np.int32)]481        )482        return new_boxes, new_scores, new_cls483 484    def _suppress_contained_lower_conf(485        self,486        boxes: np.ndarray,487        scores: np.ndarray,488        cls_ids: np.ndarray,489        target_cls: int,490        overlap: float,491    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:492        """For one class, when two boxes overlap (intersection >= `overlap` of493        the smaller box) keep the higher-confidence box and drop the other.494        Geometry is never changed -- only the redundant lower-conf box is495        removed. `overlap > 1.0` disables."""496        if overlap > 1.0:497            return boxes, scores, cls_ids498        idx = np.where(cls_ids == target_cls)[0]499        if len(idx) <= 1:500            return boxes, scores, cls_ids501 502        order = idx[np.argsort(-scores[idx])]  # highest confidence first503        remove: set[int] = set()504        for a in range(len(order)):505            i = int(order[a])506            if i in remove:507                continue508            bi = boxes[i]509            area_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))510            for b in range(a + 1, len(order)):511                j = int(order[b])512                if j in remove:513                    continue514                bj = boxes[j]515                ix1 = max(bi[0], bj[0]); iy1 = max(bi[1], bj[1])516                ix2 = min(bi[2], bj[2]); iy2 = min(bi[3], bj[3])517                inter = max(0.0, ix2 - ix1) * max(0.0, iy2 - iy1)518                if inter <= 0.0:519                    continue520                area_j = max(1e-7, float((bj[2] - bj[0]) * (bj[3] - bj[1])))521                if inter / (min(area_i, area_j) + 1e-7) >= overlap:522                    remove.add(j)  # j is the lower-confidence box (order desc)523        if not remove:524            return boxes, scores, cls_ids525        keep = np.array(526            [k not in remove for k in range(len(boxes))], dtype=bool527        )528        return boxes[keep], scores[keep], cls_ids[keep]529 530    def _merge_same_class_boxes(531        self,532        boxes: np.ndarray,533        scores: np.ndarray,534        cls_ids: np.ndarray,535    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:536        """Resolve nested / fragmented same-object detections, per class.537 538        Smoke: diffuse plumes fragment into nested boxes NMS can't collapse, so539        they are UNION-merged (smoke_merge_overlap).540        Fire: a tight hot-core box and a looser flame box are the same flame;541        keep the HIGHER-confidence one and drop the other (fire_suppress_overlap),542        which leaves geometry intact. The union-merge variant (fire_merge_overlap)543        is also available but measured worse, so it is disabled by default.544        """545        boxes, scores, cls_ids = self._merge_class_boxes(546            boxes, scores, cls_ids,547            self.class_names.index("smoke"), self.smoke_merge_overlap,548        )549        boxes, scores, cls_ids = self._merge_class_boxes(550            boxes, scores, cls_ids,551            self.class_names.index("fire"), self.fire_merge_overlap,552        )553        boxes, scores, cls_ids = self._suppress_contained_lower_conf(554            boxes, scores, cls_ids,555            self.class_names.index("fire"), self.fire_suppress_overlap,556        )557        return boxes, scores, cls_ids558 559    # Back-compat alias (older callers / tune_miner referenced this name).560    def _merge_smoke_boxes(561        self,562        boxes: np.ndarray,563        scores: np.ndarray,564        cls_ids: np.ndarray,565    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:566        return self._merge_same_class_boxes(boxes, scores, cls_ids)567 568    @staticmethod569    def _max_score_per_cluster(570        post_boxes: np.ndarray,571        post_cls: np.ndarray,572        full_boxes: np.ndarray,573        full_scores: np.ndarray,574        full_cls: np.ndarray,575        iou_thresh: float,576    ) -> np.ndarray:577        """For each kept (post-NMS) box, return the max score over the FULL578        candidate set among same-class boxes with IoU >= iou_thresh.579 580        Used after horizontal-flip TTA: a high-confidence flipped detection581        can raise the score of the corresponding original detection.582        """583        n = len(post_boxes)584        if n == 0:585            return np.empty(0, dtype=np.float32)586        full_areas = (np.maximum(0.0, full_boxes[:, 2] - full_boxes[:, 0]) *587                      np.maximum(0.0, full_boxes[:, 3] - full_boxes[:, 1]))588        out = np.empty(n, dtype=np.float32)589        for i in range(n):590            bi = post_boxes[i]591            xx1 = np.maximum(bi[0], full_boxes[:, 0])592            yy1 = np.maximum(bi[1], full_boxes[:, 1])593            xx2 = np.minimum(bi[2], full_boxes[:, 2])594            yy2 = np.minimum(bi[3], full_boxes[:, 3])595            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)596            a_i = max(0.0, float((bi[2] - bi[0]) * (bi[3] - bi[1])))597            iou = inter / (a_i + full_areas - inter + 1e-7)598            cluster = (iou >= iou_thresh) & (full_cls == post_cls[i])599            out[i] = float(np.max(full_scores[cluster])) if np.any(cluster) else 0.0600        return out601 602    def _conf_filter_mask(603        self, scores: np.ndarray, cls_ids: np.ndarray604    ) -> np.ndarray:605        """Boolean keep-mask: score >= per-class threshold, with a per-class606        rescue -- if a class has zero boxes passing, admit its top-1 candidate607        when its score >= (per-class threshold - per-class bonus)."""608        if len(scores) == 0:609            return np.zeros(0, dtype=bool)610        thr = self._conf_thres_array[cls_ids]611        keep = scores >= thr612        for c in np.unique(cls_ids):613            b = float(self._bonus_array[c])614            if b <= 0.0:615                continue616            cm = cls_ids == c617            if keep[cm].any():618                continue619            idx = np.where(cm)[0]620            top = int(idx[int(np.argmax(scores[idx]))])621            if scores[top] >= self._conf_thres_array[c] - b:622                keep[top] = True623        return keep624 625    def _filter_sane_boxes(626        self,627        boxes: np.ndarray,628        scores: np.ndarray,629        cls_ids: np.ndarray,630        orig_size: tuple[int, int],631    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:632        """Drop tiny / degenerate / image-spanning / extreme-AR boxes (FP)."""633        if len(boxes) == 0:634            return boxes, scores, cls_ids635        orig_w, orig_h = orig_size636        image_area = float(orig_w * orig_h)637        keep = []638        for i, box in enumerate(boxes):639            x1, y1, x2, y2 = box.tolist()640            bw = x2 - x1641            bh = y2 - y1642            if bw <= 0 or bh <= 0:643                continue644            if bw < self.min_side or bh < self.min_side:645                continue646            area = bw * bh647            if area < self.min_box_area:648                continue649            if area > 0.95 * image_area:650                continue651            ar = max(bw / max(bh, 1e-6), bh / max(bw, 1e-6))652            if ar > self.max_aspect_ratio:653                continue654            keep.append(i)655        if not keep:656            return (657                np.empty((0, 4), dtype=np.float32),658                np.empty((0,), dtype=np.float32),659                np.empty((0,), dtype=np.int32),660            )661        k = np.array(keep, dtype=np.intp)662        return boxes[k], scores[k], cls_ids[k]663 664    def _per_view_pipeline(665        self,666        boxes: np.ndarray,667        scores: np.ndarray,668        cls_ids: np.ndarray,669    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:670        """Per-view post-processing pipeline: per-class NMS -> cap -> cross-class dedup -> smoke merge."""671        if len(boxes) > 1:672            keep = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)673            boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]674        if len(scores) > self.max_det:675            top = np.argsort(-scores)[: self.max_det]676            boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]677        if len(boxes) > 1:678            boxes, scores, cls_ids = self._cross_class_dedup_op(679                boxes, scores, cls_ids, self.cross_iou_thresh680            )681        if len(boxes) > 1:682            boxes, scores, cls_ids = self._merge_same_class_boxes(boxes, scores, cls_ids)683        return boxes, scores, cls_ids684 685    @staticmethod686    def _roi_for_box(image: np.ndarray, box: BoundingBox) -> np.ndarray | None:687        """Clip a BoundingBox to the image and return its BGR pixel ROI."""688        h, w = image.shape[:2]689        x1 = max(0, int(math.floor(box.x1)))690        y1 = max(0, int(math.floor(box.y1)))691        x2 = min(w, int(math.ceil(box.x2)))692        y2 = min(h, int(math.ceil(box.y2)))693        if x2 <= x1 or y2 <= y1:694            return None695        roi = image[y1:y2, x1:x2]696        return roi if roi.size else None697 698    def _roi_is_near_grayscale(self, roi: np.ndarray) -> bool:699        """True if the ROI carries almost no color (validator grayscale frame).700        On such ROIs the color priors are skipped so they can't delete valid701        red/warm objects that have been stripped of color."""702        mx = roi.max(axis=2).astype(np.float32)703        mn = roi.min(axis=2).astype(np.float32)704        sat = (mx - mn) / (mx + 1e-6)705        return float(sat.mean()) < self.color_filter_min_saturation706 707    @staticmethod708    def _passes_fire_color(roi: np.ndarray) -> bool:709        """Fire is warm and/or has a bright hotspot. ROI is BGR."""710        blue = roi[:, :, 0].astype(np.float32)711        green = roi[:, :, 1].astype(np.float32)712        red = roi[:, :, 2].astype(np.float32)713        mean_r = float(np.mean(red))714        max_rgb = float(max(np.max(red), np.max(green), np.max(blue)))715        bright_frac = float(np.mean(np.max(roi, axis=2) >= 150))716        # A bright hotspot is fire-like even with little hue (also covers the717        # near-white core of an intense flame).718        if max_rgb >= 200.0 and bright_frac >= 0.01:719            return True720        warm = (red > green + 10.0) & (red > blue + 10.0)721        warm_frac = float(np.mean(warm))722        r_minus_g = mean_r - float(np.mean(green))723        if warm_frac >= 0.05 and (724            max_rgb >= 120.0 or mean_r >= 120.0 or warm_frac >= 0.15725        ):726            return True727        if bright_frac >= 0.12 and r_minus_g >= 2.0:728            return True729        return False730 731    @staticmethod732    def _passes_fire_ext_red_color(roi: np.ndarray) -> bool:733        """Fire extinguishers are red. ROI is BGR. Lenient: only clearly734        cool/green/blue or very dark regions fail."""735        blue = roi[:, :, 0].astype(np.float32)736        green = roi[:, :, 1].astype(np.float32)737        red = roi[:, :, 2].astype(np.float32)738        red_dom = float(np.mean((red > green + 10.0) & (red > blue + 10.0)))739        if red_dom >= 0.03:740            return True741        if (float(np.mean(red)) - float(np.mean(green))) >= 0.0 and \742                float(np.mean(red)) >= 50.0:743            return True744        return False745 746    def _remove_edge_low_conf(747        self, results: list[BoundingBox], orig_size: tuple[int, int]748    ) -> list[BoundingBox]:749        """Drop border-hugging boxes in the low-confidence band."""750        if (751            not self.use_edge_filter752            or self.edge_filter_max_conf <= 0.0753            or not results754        ):755            return results756        w, h = orig_size757        tol = self.edge_tol758        out: list[BoundingBox] = []759        for b in results:760            on_edge = (761                b.x1 <= tol762                or b.y1 <= tol763                or b.x2 >= w - 1 - tol764                or b.y2 >= h - 1 - tol765            )766            if on_edge and b.conf <= self.edge_filter_max_conf:767                continue768            out.append(b)769        return out770 771    def _views_corroborated(772        self,773        post_boxes: np.ndarray,774        post_cls: np.ndarray,775        full_boxes: np.ndarray,776        full_cls: np.ndarray,777        full_views: np.ndarray,778        iou_thresh: float,779    ) -> np.ndarray:780        """For each post-NMS box, True if same-class detections from >= 2781        distinct TTA views overlap it (IoU >= iou_thresh) in the full union."""782        n = len(post_boxes)783        if n == 0:784            return np.zeros(0, dtype=bool)785        full_areas = (np.maximum(0.0, full_boxes[:, 2] - full_boxes[:, 0]) *786                      np.maximum(0.0, full_boxes[:, 3] - full_boxes[:, 1]))787        out = np.zeros(n, dtype=bool)788        for i in range(n):789            bi = post_boxes[i]790            xx1 = np.maximum(bi[0], full_boxes[:, 0])791            yy1 = np.maximum(bi[1], full_boxes[:, 1])792            xx2 = np.minimum(bi[2], full_boxes[:, 2])793            yy2 = np.minimum(bi[3], full_boxes[:, 3])794            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)795            a_i = max(0.0, float((bi[2] - bi[0]) * (bi[3] - bi[1])))796            iou = inter / (a_i + full_areas - inter + 1e-7)797            mask = (iou >= iou_thresh) & (full_cls == post_cls[i])798            if np.any(mask):799                out[i] = len(np.unique(full_views[mask])) >= 2800        return out801 802    def _filter_low_conf_by_color(803        self, image: np.ndarray, results: list[BoundingBox]804    ) -> list[BoundingBox]:805        """Drop borderline fire / extinguisher detections whose pixels clearly806        contradict the class's expected color. No-op on near-grayscale ROIs and807        on detections above the per-class color-filter conf gate."""808        if not results:809            return results810        cls_fire = self.class_names.index("fire")811        cls_ext = self.class_names.index("fire extinguisher")812        out: list[BoundingBox] = []813        for box in results:814            check_fire = (815                box.cls_id == cls_fire816                and box.conf <= self.fire_color_filter_max_conf817            )818            check_ext = (819                box.cls_id == cls_ext820                and box.conf <= self.fire_ext_color_filter_max_conf821            )822            if not check_fire and not check_ext:823                out.append(box)824                continue825            roi = self._roi_for_box(image, box)826            if roi is None or self._roi_is_near_grayscale(roi):827                out.append(box)828                continue829            if check_fire and not self._passes_fire_color(roi):830                continue831            if check_ext and not self._passes_fire_ext_red_color(roi):832                continue833            out.append(box)834        return out835 836    @staticmethod837    def _build_results(838        boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray839    ) -> list[BoundingBox]:840        results: list[BoundingBox] = []841        for box, conf, cls_id in zip(boxes, scores, cls_ids):842            x1, y1, x2, y2 = box.tolist()843            if x2 <= x1 or y2 <= y1:844                continue845            results.append(846                BoundingBox(847                    x1=int(math.floor(x1)),848                    y1=int(math.floor(y1)),849                    x2=int(math.ceil(x2)),850                    y2=int(math.ceil(y2)),851                    cls_id=int(cls_id),852                    conf=float(conf),853                )854            )855        return results856 857    def _decode_final_dets(858        self,859        preds: np.ndarray,860        ratio: float,861        pad: tuple[float, float],862        orig_size: tuple[int, int],863    ) -> list[BoundingBox]:864        """Final-detection output path: rows shaped [x1, y1, x2, y2, conf, cls_id]."""865        if preds.ndim == 3 and preds.shape[0] == 1:866            preds = preds[0]867        if preds.ndim != 2 or preds.shape[1] < 6:868            raise ValueError(f"Unexpected ONNX final-det output shape: {preds.shape}")869 870        boxes = preds[:, :4].astype(np.float32)871        scores = preds[:, 4].astype(np.float32)872        cls_ids = preds[:, 5].astype(np.int32)873        cls_ids = self.cls_remap[cls_ids]874 875        keep = self._conf_filter_mask(scores, cls_ids)876        boxes = boxes[keep]877        scores = scores[keep]878        cls_ids = cls_ids[keep]879        if len(boxes) == 0:880            return []881 882        pad_w, pad_h = pad883        boxes[:, [0, 2]] -= pad_w884        boxes[:, [1, 3]] -= pad_h885        boxes /= ratio886        boxes = self._clip_boxes(boxes, orig_size)887 888        boxes, scores, cls_ids = self._filter_sane_boxes(889            boxes, scores, cls_ids, orig_size890        )891        if len(boxes) == 0:892            return []893 894        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)895        return self._build_results(boxes, scores, cls_ids)896 897    def _decode_raw_yolo(898        self,899        preds: np.ndarray,900        ratio: float,901        pad: tuple[float, float],902        orig_size: tuple[int, int],903    ) -> list[BoundingBox]:904        """Fallback raw-YOLO output path: per-anchor class logits."""905        if preds.ndim != 3 or preds.shape[0] != 1:906            raise ValueError(f"Unexpected raw ONNX output shape: {preds.shape}")907        preds = preds[0]908        if preds.shape[0] <= 16 and preds.shape[1] > preds.shape[0]:909            preds = preds.T910        if preds.ndim != 2 or preds.shape[1] < 5:911            raise ValueError(f"Unexpected raw output shape: {preds.shape}")912 913        boxes_xywh = preds[:, :4].astype(np.float32)914        cls_part = preds[:, 4:].astype(np.float32)915        if cls_part.shape[1] == 1:916            scores = cls_part[:, 0]917            cls_ids = np.zeros(len(scores), dtype=np.int32)918        else:919            cls_ids = np.argmax(cls_part, axis=1).astype(np.int32)920            scores = cls_part[np.arange(len(cls_part)), cls_ids]921        cls_ids = self.cls_remap[cls_ids]922 923        keep = self._conf_filter_mask(scores, cls_ids)924        boxes_xywh = boxes_xywh[keep]925        scores = scores[keep]926        cls_ids = cls_ids[keep]927        if len(boxes_xywh) == 0:928            return []929        boxes = self._xywh_to_xyxy(boxes_xywh)930 931        pad_w, pad_h = pad932        boxes[:, [0, 2]] -= pad_w933        boxes[:, [1, 3]] -= pad_h934        boxes /= ratio935        boxes = self._clip_boxes(boxes, orig_size)936 937        boxes, scores, cls_ids = self._filter_sane_boxes(938            boxes, scores, cls_ids, orig_size939        )940        if len(boxes) == 0:941            return []942 943        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)944        return self._build_results(boxes, scores, cls_ids)945 946    def _postprocess(947        self,948        output: np.ndarray,949        ratio: float,950        pad: tuple[float, float],951        orig_size: tuple[int, int],952    ) -> list[BoundingBox]:953        if output.ndim == 2 and output.shape[1] >= 6:954            return self._decode_final_dets(output, ratio, pad, orig_size)955        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 6:956            return self._decode_final_dets(output, ratio, pad, orig_size)957        return self._decode_raw_yolo(output, ratio, pad, orig_size)958 959    def _predict_single(self, image: np.ndarray) -> list[BoundingBox]:960        if image is None:961            raise ValueError("Input image is None")962        if not isinstance(image, np.ndarray):963            raise TypeError(f"Input is not numpy array: {type(image)}")964        if image.ndim != 3:965            raise ValueError(f"Expected HWC image, got shape={image.shape}")966        if image.shape[0] <= 0 or image.shape[1] <= 0:967            raise ValueError(f"Invalid image shape={image.shape}")968        if image.shape[2] != 3:969            raise ValueError(f"Expected 3 channels, got shape={image.shape}")970        if image.dtype != np.uint8:971            image = image.astype(np.uint8)972 973        input_tensor, ratio, pad, orig_size = self._preprocess(image)974        expected = (1, 3, self.input_height, self.input_width)975        if input_tensor.shape != expected:976            raise ValueError(977                f"Bad input tensor shape={input_tensor.shape}, expected={expected}"978            )979 980        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})981        return self._postprocess(outputs[0], ratio, pad, orig_size)982 983    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:984        """Horizontal-flip TTA.985 986        Strategy:987          1. Predict on original and on flipped image.988          2. Map flipped boxes back to original coordinates.989          3. Per-class hard NMS on the union.990          4. For each kept box, compute the max same-class score across the991             FULL union (not just the post-NMS subset) -- this lets a high-992             confidence flipped detection raise a borderline original one.993          5. Cross-class dedup to suppress same-physical-object multi-class.994          6. Smoke merge: overlapping / nested smoke boxes collapse into995             their union (one box per smoke object).996        """997        boxes_orig = self._predict_single(image)998        flipped = cv2.flip(image, 1)999        boxes_flip = self._predict_single(flipped)1000        w = image.shape[1]1001        boxes_flip = [1002            BoundingBox(1003                x1=w - b.x2, y1=b.y1, x2=w - b.x1, y2=b.y2,1004                cls_id=b.cls_id, conf=b.conf,1005            )1006            for b in boxes_flip1007        ]1008        all_boxes = boxes_orig + boxes_flip1009        if not all_boxes:1010            return []1011 1012        coords = np.array(1013            [[b.x1, b.y1, b.x2, b.y2] for b in all_boxes], dtype=np.float321014        )1015        scores = np.array([b.conf for b in all_boxes], dtype=np.float32)1016        cls_ids = np.array([b.cls_id for b in all_boxes], dtype=np.int32)1017        # view_id 0 = original, 1 = horizontal flip (mapped back to orig coords)1018        view_ids = np.array(1019            [0] * len(boxes_orig) + [1] * len(boxes_flip), dtype=np.int321020        )1021 1022        hard_keep = self._per_class_hard_nms(coords, scores, cls_ids, self.iou_thres)1023        if len(hard_keep) == 0:1024            return []1025        if len(hard_keep) > self.max_det:1026            top = np.argsort(-scores[hard_keep])[: self.max_det]1027            hard_keep = hard_keep[top]1028 1029        boosted = self._max_score_per_cluster(1030            coords[hard_keep], cls_ids[hard_keep],1031            coords, scores, cls_ids, self.iou_thres,1032        )1033 1034        kept_coords = coords[hard_keep]1035        kept_cls = cls_ids[hard_keep]1036 1037        # Optional: drop low-conf detections seen in only one TTA view.1038        if (1039            self.use_tta_view_filter1040            and self.tta_view_filter_max_conf > 0.01041            and len(kept_coords) > 01042        ):1043            corrob = self._views_corroborated(1044                kept_coords, kept_cls, coords, cls_ids, view_ids,1045                self.tta_view_iou_thresh,1046            )1047            keep = ~((boosted <= self.tta_view_filter_max_conf) & (~corrob))1048            kept_coords = kept_coords[keep]1049            boosted = boosted[keep]1050            kept_cls = kept_cls[keep]1051 1052        if len(kept_coords) > 1:1053            kept_coords, boosted, kept_cls = self._cross_class_dedup_op(1054                kept_coords, boosted, kept_cls, self.cross_iou_thresh1055            )1056        if len(kept_coords) > 1:1057            kept_coords, boosted, kept_cls = self._merge_same_class_boxes(1058                kept_coords, boosted, kept_cls1059            )1060 1061        return [1062            BoundingBox(1063                x1=int(math.floor(kept_coords[j, 0])),1064                y1=int(math.floor(kept_coords[j, 1])),1065                x2=int(math.ceil(kept_coords[j, 2])),1066                y2=int(math.ceil(kept_coords[j, 3])),1067                cls_id=int(kept_cls[j]),1068                conf=float(boosted[j]),1069            )1070            for j in range(len(kept_coords))1071        ]1072 1073    def predict_batch(1074        self,1075        batch_images: list[ndarray],1076        offset: int,1077        n_keypoints: int,1078    ) -> list[TVFrameResult]:1079        results: list[TVFrameResult] = []1080        for frame_number_in_batch, image in enumerate(batch_images):1081            try:1082                if self.use_tta:1083                    boxes = self._predict_tta(image)1084                else:1085                    boxes = self._predict_single(image)1086                # Color-prior + edge FP filters on the merged result, in1087                # original-image coords. Single insertion point so they run once1088                # per frame for both the TTA and non-TTA paths.1089                if isinstance(image, np.ndarray) and image.ndim == 3:1090                    boxes = self._filter_low_conf_by_color(image, boxes)1091                    boxes = self._remove_edge_low_conf(1092                        boxes, (image.shape[1], image.shape[0])1093                    )1094            except Exception as e:1095                print(1096                    f"⚠️ Inference failed for frame "1097                    f"{offset + frame_number_in_batch}: {e}"1098                )1099                boxes = []1100            results.append(1101                TVFrameResult(1102                    frame_id=offset + frame_number_in_batch,1103                    boxes=boxes,1104                    keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],1105                )1106            )1107        return results1108 1109    def predict_image(self, image: ndarray) -> list[BoundingBox]:1110        """Run detection on a single BGR image."""1111        if self.use_tta:1112            boxes = self._predict_tta(image)1113        else:1114            boxes = self._predict_single(image)1115        if isinstance(image, np.ndarray) and image.ndim == 3:1116            boxes = self._filter_low_conf_by_color(image, boxes)1117            boxes = self._remove_edge_low_conf(boxes, (image.shape[1], image.shape[0]))1118        return boxes1119