"""TurboVision crime-detection miner. YOLO11s @ 1280x1280, 6-class detection (balaclava, bat, glove, graffiti, hoodie, spray paint), ONNX with end-to-end NMS baked in. Output of weights.onnx: [1, 300, 6] = x1, y1, x2, y2, conf, cls (post-NMS). Inference pipeline: 1) Primary forward pass on the full image. 2) Hflip TTA: forward on horizontally-flipped image, transform boxes back. 3) Per-class hard-NMS to merge primary + flip outputs. 4) Cross-class IoU dedup (suppresses same physical object getting two class labels). 5) Consensus-confidence boost: when both views agree on a cluster, take max score. 6) Sanity filter (min size, aspect ratio). Class taxonomy (must match the validator manifest's `objects` list for this element): 0 balaclava 1 bat 2 glove 3 graffiti 4 hoodie 5 spray paint """ 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: def __init__(self, path_hf_repo: Path) -> None: model_path = path_hf_repo / "weights.onnx" # Validator manifest order (from spec.json `objects`): # 0=balaclava 1=hoodie 2=glove 3=bat 4="spray paint" 5=graffiti # v5 weights.onnx was trained with this exact order, so cls_remap is identity. cn_path = model_path.with_name("class_names.txt") if cn_path.is_file(): self.class_names = [ ln.strip() for ln in cn_path.read_text(encoding="utf-8").splitlines() if ln.strip() and not ln.strip().startswith("#") ] else: self.class_names = ["balaclava", "hoodie", "glove", "bat", "spray paint", "graffiti"] self.cls_remap = np.arange(len(self.class_names), dtype=np.int32) print("ORT version:", ort.__version__) try: ort.preload_dlls() print("✅ onnxruntime.preload_dlls() success") except Exception as e: print(f"⚠️ preload_dlls failed: {e}") print("ORT available providers BEFORE session:", ort.get_available_providers()) sess_options = ort.SessionOptions() sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL try: self.session = ort.InferenceSession( str(model_path), sess_options=sess_options, providers=["CUDAExecutionProvider", "CPUExecutionProvider"], ) print("✅ Created ORT session with preferred CUDA provider list") except Exception as e: print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}") self.session = ort.InferenceSession( str(model_path), sess_options=sess_options, providers=["CPUExecutionProvider"], ) print("ORT session providers:", self.session.get_providers()) inp = self.session.get_inputs()[0] self.input_name = inp.name self.output_names = [o.name for o in self.session.get_outputs()] self.input_shape = inp.shape self.input_dtype = np.float16 if "float16" in inp.type else np.float32 self.input_height = self._safe_dim(self.input_shape[2], default=1280) self.input_width = self._safe_dim(self.input_shape[3], default=1280) # Tuning matched to alfred's deployed model — bias toward precision to dodge # the false_positive pillar penalty (validator weights FP heavily on this element). # v13 sweet spot on starter (true GT): uniform conf=0.50. # Tuning per-class on 7 images overfits — leave-one-out CV showed it # collapsed to 0.314 on held-out shards. Uniform 0.50 is robust. self.conf_thres = 0.50 self.conf_thres_per_class = np.array([0.50] * 6, dtype=np.float32) self.iou_thres = 0.4 self.cross_iou_thresh = 0.7 self.max_det = 100 self.use_tta = False # Sanity filter — reject obviously bad boxes self.min_box_area = 14 * 14 self.min_side = 8 self.max_aspect_ratio = 8.0 self.max_box_area_ratio = 0.95 print(f"✅ ONNX loaded: {model_path}") print(f"✅ providers: {self.session.get_providers()}") print(f"✅ input: name={self.input_name}, shape={self.input_shape}, dtype={self.input_dtype}") print(f"✅ classes: {self.class_names}") print(f"✅ config: conf={self.conf_thres}, iou={self.iou_thres}, " f"cross_iou={self.cross_iou_thresh}, TTA={self.use_tta}") def __repr__(self) -> str: return ( f"ONNXRuntime(session={type(self.session).__name__}, " f"providers={self.session.get_providers()})" ) @staticmethod def _safe_dim(value, default: int) -> int: return value if isinstance(value, int) and value > 0 else default def _letterbox( self, image: ndarray, new_shape: tuple[int, int], color=(114, 114, 114), ) -> tuple[ndarray, float, tuple[float, float]]: h, w = image.shape[:2] new_w, new_h = new_shape ratio = min(new_w / w, new_h / h) resized_w = int(round(w * ratio)) resized_h = int(round(h * ratio)) if (resized_w, resized_h) != (w, h): interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR image = cv2.resize(image, (resized_w, resized_h), interpolation=interp) dw = (new_w - resized_w) / 2.0 dh = (new_h - resized_h) / 2.0 left = int(round(dw - 0.1)) right = int(round(dw + 0.1)) top = int(round(dh - 0.1)) bottom = int(round(dh + 0.1)) padded = cv2.copyMakeBorder( image, top, bottom, left, right, borderType=cv2.BORDER_CONSTANT, value=color, ) return padded, ratio, (dw, dh) def _preprocess(self, image: ndarray): orig_h, orig_w = image.shape[:2] img, ratio, pad = self._letterbox(image, (self.input_width, self.input_height)) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) img = img.astype(self.input_dtype) / 255.0 img = np.transpose(img, (2, 0, 1))[None, ...] img = np.ascontiguousarray(img) return img, ratio, pad, (orig_w, orig_h) @staticmethod def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray: w, h = image_size boxes[:, 0] = np.clip(boxes[:, 0], 0, w - 1) boxes[:, 1] = np.clip(boxes[:, 1], 0, h - 1) boxes[:, 2] = np.clip(boxes[:, 2], 0, w - 1) boxes[:, 3] = np.clip(boxes[:, 3], 0, h - 1) return boxes def _filter_sane_boxes( self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray, orig_size: tuple[int, int], ): 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 > self.max_box_area_ratio * 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] @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) boxes = np.asarray(boxes, dtype=np.float32) scores = np.asarray(scores, dtype=np.float32) order = np.argsort(scores)[::-1] keep: list[int] = [] suppressed = np.zeros(N, dtype=bool) for i in range(N): idx = order[i] if suppressed[idx]: continue keep.append(int(idx)) bi = boxes[idx] for k in range(i + 1, N): jdx = order[k] if suppressed[jdx]: continue bj = boxes[jdx] xx1 = max(bi[0], bj[0]) yy1 = max(bi[1], bj[1]) xx2 = min(bi[2], bj[2]) yy2 = min(bi[3], bj[3]) inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1) area_i = (bi[2] - bi[0]) * (bi[3] - bi[1]) area_j = (bj[2] - bj[0]) * (bj[3] - bj[1]) iou = inter / (area_i + area_j - inter + 1e-7) if iou > iou_thresh: suppressed[jdx] = True return np.array(keep, dtype=np.intp) def _per_class_hard_nms( self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray, iou_thresh: float, ) -> np.ndarray: if len(boxes) == 0: return np.array([], dtype=np.intp) all_keep: list[int] = [] for c in np.unique(cls_ids): mask = cls_ids == c indices = np.where(mask)[0] keep = self._hard_nms(boxes[mask], scores[mask], iou_thresh) all_keep.extend(indices[keep].tolist()) all_keep.sort() return np.array(all_keep, dtype=np.intp) @staticmethod def _cross_class_dedup( boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray, iou_thresh: float, ) -> tuple[np.ndarray, np.ndarray, np.ndarray]: 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] ) # Keep larger boxes first, then higher score. order = np.lexsort((-scores, -areas)) 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) area_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1]))) union = area_i + areas - inter + 1e-7 iou = inter / union dup = iou > iou_thresh dup[i] = False suppressed |= dup keep_idx = np.array(keep, dtype=np.intp) return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx] @staticmethod def _max_score_per_cluster( coords: np.ndarray, scores: np.ndarray, keep_indices: np.ndarray, iou_thresh: float, ) -> np.ndarray: n_keep = len(keep_indices) if n_keep == 0: return np.array([], dtype=np.float32) coords = np.asarray(coords, dtype=np.float32) scores = np.asarray(scores, dtype=np.float32) out = np.empty(n_keep, dtype=np.float32) for i in range(n_keep): idx = keep_indices[i] bi = coords[idx] xx1 = np.maximum(bi[0], coords[:, 0]) yy1 = np.maximum(bi[1], coords[:, 1]) xx2 = np.minimum(bi[2], coords[:, 2]) yy2 = np.minimum(bi[3], coords[:, 3]) inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1) area_i = (bi[2] - bi[0]) * (bi[3] - bi[1]) areas_j = (coords[:, 2] - coords[:, 0]) * (coords[:, 3] - coords[:, 1]) iou = inter / (area_i + areas_j - inter + 1e-7) in_cluster = iou >= iou_thresh out[i] = float(np.max(scores[in_cluster])) return out def _decode_raw_dets( self, preds: np.ndarray, ratio: float, pad: tuple[float, float], orig_size: tuple[int, int], *, apply_conf_thresh: bool = True, ) -> tuple[np.ndarray, np.ndarray, np.ndarray]: """Decode end2end NMS output and return (boxes, scores, cls_ids) in original image coordinates, after conf-threshold + remap + letterbox-reverse + sanity. When apply_conf_thresh=False, the conf-threshold filter is skipped (used for the no-detection fallback path: take the single top-conf raw box).""" 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 output shape: {preds.shape}") boxes = preds[:, :4].astype(np.float32) scores = preds[:, 4].astype(np.float32) cls_ids = preds[:, 5].astype(np.int32) valid = (cls_ids >= 0) & (cls_ids < len(self.cls_remap)) & (scores > 0) boxes, scores, cls_ids = boxes[valid], scores[valid], cls_ids[valid] cls_ids = self.cls_remap[cls_ids] if apply_conf_thresh: # Per-class threshold: each box compared against its own class's threshold cls_thresh = np.full(len(scores), self.conf_thres, dtype=np.float32) valid_cls = (cls_ids >= 0) & (cls_ids < len(self.conf_thres_per_class)) cls_thresh[valid_cls] = self.conf_thres_per_class[cls_ids[valid_cls]] keep = scores >= cls_thresh boxes = boxes[keep] scores = scores[keep] cls_ids = cls_ids[keep] if len(boxes) == 0: return ( np.empty((0, 4), dtype=np.float32), np.empty((0,), dtype=np.float32), np.empty((0,), dtype=np.int32), ) pad_w, pad_h = pad orig_w, orig_h = orig_size boxes[:, [0, 2]] -= pad_w boxes[:, [1, 3]] -= pad_h boxes /= ratio boxes = self._clip_boxes(boxes, (orig_w, orig_h)) boxes, scores, cls_ids = self._filter_sane_boxes(boxes, scores, cls_ids, orig_size) return boxes, scores, cls_ids def _forward( self, image: np.ndarray ) -> tuple[np.ndarray, np.ndarray, np.ndarray]: x, ratio, pad, orig_size = self._preprocess(image) out = self.session.run(self.output_names, {self.input_name: x})[0] return self._decode_raw_dets(out, ratio, pad, orig_size) def _forward_with_fallback( self, image: np.ndarray ) -> tuple[ tuple[np.ndarray, np.ndarray, np.ndarray], tuple[np.ndarray, np.ndarray, np.ndarray], ]: """Run ONNX once, decode twice: (filtered @ conf_thres, all-survived sanity).""" x, ratio, pad, orig_size = self._preprocess(image) out = self.session.run(self.output_names, {self.input_name: x})[0] primary = self._decode_raw_dets(out, ratio, pad, orig_size, apply_conf_thresh=True) fallback = self._decode_raw_dets(out, ratio, pad, orig_size, apply_conf_thresh=False) return primary, fallback def _predict_single(self, image: np.ndarray) -> list[BoundingBox]: (boxes, scores, cls_ids), (fb_b, fb_s, fb_c) = self._forward_with_fallback(image) ih, iw = image.shape[:2] if len(boxes) > 0: return self._build_results(boxes, scores, cls_ids, image_size=(iw, ih)) # FALLBACK: nothing passed conf_thres — return single top-conf box # (any class, any conf > 0) so the validator's mAP isn't a hard zero. if len(fb_b) == 0: return [] i = int(np.argmax(fb_s)) return self._build_results( fb_b[i:i + 1], fb_s[i:i + 1], fb_c[i:i + 1], image_size=(iw, ih) ) def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]: """Hflip TTA: merge primary + flipped via per-class hard-NMS, then cross-class dedup, with consensus-confidence boost.""" ow = image.shape[1] b1, s1, c1 = self._forward(image) flipped = cv2.flip(image, 1) b2, s2, c2 = self._forward(flipped) if len(b2): x1f = ow - b2[:, 2] x2f = ow - b2[:, 0] b2 = np.stack([x1f, b2[:, 1], x2f, b2[:, 3]], axis=1) if len(b1) == 0 and len(b2) == 0: return [] boxes = np.concatenate([b1, b2], axis=0) if len(b2) else b1 scores = np.concatenate([s1, s2], axis=0) if len(b2) else s1 cls_ids = np.concatenate([c1, c2], axis=0) if len(b2) else c1 keep = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres) if len(keep) == 0: return [] keep = keep[: self.max_det] # Consensus-confidence boost: cluster by IoU and take max score. boosted = self._max_score_per_cluster(boxes, scores, keep, self.iou_thres) boxes = boxes[keep] cls_ids = cls_ids[keep] scores = boosted boxes, scores, cls_ids = self._cross_class_dedup( boxes, scores, cls_ids, self.cross_iou_thresh ) if len(boxes) == 0: return [] ih, iw = image.shape[:2] return self._build_results(boxes, scores, cls_ids, image_size=(iw, ih)) def _filter_balaclava_geometry( self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray, image_size: tuple[int, int] | None = None, ) -> tuple[np.ndarray, np.ndarray, np.ndarray]: # Real-balaclava prior (from 43 manual GT labels): # aspect ratio max(w/h, h/w): p5=1.11, median=1.33, p99=1.71 # rel area % of image: p1=0.041, p5=0.070, p10=0.087 # FP balaclavas frequently violate these (very thin/wide boxes from # face-fragment matches, or tiny ~0.01%-area boxes from texture noise). BALACLAVA = 0 ASPECT_MAX = 1.8 # above p99 of real REL_AREA_MIN = 0.0004 # below p1 of real (0.04%) if len(boxes) == 0: return boxes, scores, cls_ids is_bal = cls_ids == BALACLAVA if not is_bal.any(): return boxes, scores, cls_ids keep = np.ones(len(boxes), dtype=bool) if image_size is not None: iw, ih = image_size img_area = max(1.0, iw * ih) else: img_area = None for i in np.where(is_bal)[0]: x1, y1, x2, y2 = boxes[i] bw = max(1.0, x2 - x1) bh = max(1.0, y2 - y1) aspect = max(bw / bh, bh / bw) if aspect > ASPECT_MAX: keep[i] = False continue if img_area is not None: rel = (bw * bh) / img_area if rel < REL_AREA_MIN: keep[i] = False return boxes[keep], scores[keep], cls_ids[keep] def _suppress_balaclava_under_hoodie( self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray, ) -> tuple[np.ndarray, np.ndarray, np.ndarray]: # Validator rule: "balaclavas worn under a hoodie hood are IGNORED # (a hoodie includes the jacket and its hood)". A small balaclava # box can sit fully inside a much larger hoodie box — IoU between # them stays low (intersection / large union), but containment # (intersection / balaclava_area) is ~1.0. So drop any balaclava # whose containment by any hoodie box is >= COVER_THRESH. BALACLAVA, HOODIE = 0, 1 COVER_THRESH = 0.5 if len(boxes) == 0: return boxes, scores, cls_ids is_hood = cls_ids == HOODIE is_bal = cls_ids == BALACLAVA if not is_hood.any() or not is_bal.any(): return boxes, scores, cls_ids hood_boxes = boxes[is_hood] keep = np.ones(len(boxes), dtype=bool) for i in np.where(is_bal)[0]: bx1, by1, bx2, by2 = boxes[i] bal_area = max(1.0, (bx2 - bx1) * (by2 - by1)) ix1 = np.maximum(bx1, hood_boxes[:, 0]) iy1 = np.maximum(by1, hood_boxes[:, 1]) ix2 = np.minimum(bx2, hood_boxes[:, 2]) iy2 = np.minimum(by2, hood_boxes[:, 3]) iw = np.clip(ix2 - ix1, 0.0, None) ih = np.clip(iy2 - iy1, 0.0, None) inter = iw * ih cover = inter / bal_area if (cover >= COVER_THRESH).any(): keep[i] = False return boxes[keep], scores[keep], cls_ids[keep] def _build_results( self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray, image_size: tuple[int, int] | None = None, ) -> list[BoundingBox]: boxes, scores, cls_ids = self._filter_balaclava_geometry( boxes, scores, cls_ids, image_size ) boxes, scores, cls_ids = self._suppress_balaclava_under_hoodie( boxes, scores, cls_ids ) results: list[BoundingBox] = [] for box, conf, cls_id in zip(boxes, scores, cls_ids): x1, y1, x2, y2 = box.tolist() if x2 <= x1 or y2 <= y1: continue results.append( BoundingBox( x1=int(math.floor(x1)), y1=int(math.floor(y1)), x2=int(math.ceil(x2)), y2=int(math.ceil(y2)), cls_id=int(cls_id), conf=float(conf), ) ) return results 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): if image is None or not isinstance(image, np.ndarray) or image.ndim != 3: results.append( TVFrameResult( frame_id=offset + frame_number_in_batch, boxes=[], keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))], ) ) continue if image.dtype != np.uint8: image = image.astype(np.uint8) try: if self.use_tta: boxes = self._predict_tta(image) else: boxes = self._predict_single(image) except Exception as e: print(f"⚠️ Inference failed for frame {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