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