scorevision: push artifact

miner.py

@@ -1,434 +1,637 @@
-"""Plate-detection miner — v2.0 "hermestech + tile s<3".
-
-Base weights: hermestech00/numberplate0 (YOLO26s retrained, fp16, ~19 MB).
-
-Inference pipeline:
-  1) Full-image primary pass with alfred001 tuning
-     (conf=0.22, iou=0.41, sigma=0.685, soft-NMS + hflip TTA).
-  2) If the primary returned fewer than 3 boxes, run a 2x2
-     overlapping tile pass (tile_conf=0.40) with novelty-merge.
-"""
-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]]
-
-
-SIZE = 1280
-
-
-class Miner:
-    def __init__(self, path_hf_repo: Path) -> None:
-        model_path = path_hf_repo / "weights.onnx"
-        cn_path = model_path.with_name("class_names.txt")
-        if cn_path.is_file():
-            lines = cn_path.read_text(encoding="utf-8").splitlines()
-            self.class_names = [
-                ln.strip()
-                for ln in lines
-                if ln.strip() and not ln.strip().startswith("#")
-            ]
-        else:
-            self.class_names = ["numberplate"]
-        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())
-
-        try:
-            import torch
-            if torch.cuda.is_available():
-                print(f"GPU: {torch.cuda.get_device_name(0)}")
-                print(f"GPU memory: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB")
-            else:
-                print("GPU: CUDA not available via torch")
-        except Exception as e:
-            print(f"GPU detection failed: {e}")
-
-        sess_options = ort.SessionOptions()
-        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-
-        try:
-            self.session = ort.InferenceSession(
-                str(model_path),
-                sess_options=sess_options,
-                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
-            )
-            print("Created ORT session with preferred CUDA provider list")
-        except Exception as e:
-            print(f"CUDA session creation failed, falling back to CPU: {e}")
-            self.session = ort.InferenceSession(
-                str(model_path),
-                sess_options=sess_options,
-                providers=["CPUExecutionProvider"],
-            )
-
-        print("ORT session providers:", self.session.get_providers())
-
-        for inp in self.session.get_inputs():
-            print("INPUT:", inp.name, inp.shape, inp.type)
-        for out in self.session.get_outputs():
-            print("OUTPUT:", out.name, out.shape, out.type)
-
-        self.input_name = self.session.get_inputs()[0].name
-        self.output_names = [o.name for o 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=SIZE)
-        self.input_width = self._safe_dim(self.input_shape[3], default=SIZE)
-
-        # Primary pass: alfred001 tuning (optimized for hermestech weights)
-        self.conf_thres = 0.22
-        self.iou_thres = 0.41
-        self.sigma = 0.685
-        self.max_det = 300
-
-        # Conditional tile-pass (trimmed for latency: no hflip, tighter sparse)
-        self.sparse_threshold = 3       # fire tiles only if primary returns < this
-        self.tile_conf = 0.40
-        self.tile_overlap = 0.20
-        self.novelty_iou = 0.10
-        self.final_max_det = 22
-        self.tile_use_hflip = False     # skip hflip tile pass to save ~4 forwards
-
-        self.use_tta = True
-
-        print(f"ONNX model loaded from: {model_path}")
-        print(f"ONNX providers: {self.session.get_providers()}")
-        print(f"ONNX input: name={self.input_name}, shape={self.input_shape}")
-
-    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
-
-    # ---------- image preprocessing ----------
-    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):
-        img, ratio, pad = self._letterbox(image, (self.input_width, self.input_height))
-        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
-        img = np.transpose(img, (2, 0, 1))[None, ...]
-        return np.ascontiguousarray(img, dtype=np.float32), ratio, pad
-
-    @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
-
-    # ---------- NMS primitives ----------
-    @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)
-        keep: list[int] = []
-        while len(order):
-            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)
-            area_i = (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
-            area_r = (boxes[rest, 2] - boxes[rest, 0]) * (boxes[rest, 3] - boxes[rest, 1])
-            iou = inter / (area_i + area_r - inter + 1e-7)
-            order = rest[iou <= iou_thresh]
-        return np.array(keep, dtype=np.intp)
-
-    def _soft_nms(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        sigma: float,
-        score_thresh: float = 0.01,
-    ) -> tuple[np.ndarray, np.ndarray]:
-        N = len(boxes)
-        if N == 0:
-            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
-        boxes = boxes.astype(np.float32, copy=True)
-        scores = scores.astype(np.float32, copy=True)
-        order = np.arange(N)
-        for i in range(N):
-            max_pos = i + int(np.argmax(scores[i:]))
-            boxes[[i, max_pos]] = boxes[[max_pos, i]]
-            scores[[i, max_pos]] = scores[[max_pos, i]]
-            order[[i, max_pos]] = order[[max_pos, i]]
-            if i + 1 >= N:
-                break
-            xx1 = np.maximum(boxes[i, 0], boxes[i + 1:, 0])
-            yy1 = np.maximum(boxes[i, 1], boxes[i + 1:, 1])
-            xx2 = np.minimum(boxes[i, 2], boxes[i + 1:, 2])
-            yy2 = np.minimum(boxes[i, 3], boxes[i + 1:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            area_i = float(
-                (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
-            )
-            areas_j = (
-                np.maximum(0.0, boxes[i + 1:, 2] - boxes[i + 1:, 0])
-                * np.maximum(0.0, boxes[i + 1:, 3] - boxes[i + 1:, 1])
-            )
-            iou = inter / (area_i + areas_j - inter + 1e-7)
-            scores[i + 1:] *= np.exp(-(iou ** 2) / sigma)
-        mask = scores > score_thresh
-        return order[mask], scores[mask]
-
-    @staticmethod
-    def _box_iou_one_to_many(box: np.ndarray, boxes: np.ndarray) -> np.ndarray:
-        if len(boxes) == 0:
-            return np.zeros(0, dtype=np.float32)
-        xx1 = np.maximum(box[0], boxes[:, 0])
-        yy1 = np.maximum(box[1], boxes[:, 1])
-        xx2 = np.minimum(box[2], boxes[:, 2])
-        yy2 = np.minimum(box[3], boxes[:, 3])
-        inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-        area_a = max(0.0, (box[2] - box[0]) * (box[3] - box[1]))
-        area_b = np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) * np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
-        return inter / (area_a + area_b - inter + 1e-7)
-
-    # ---------- raw-dets helper ----------
-    def _raw_dets(self, image: ndarray, conf: float) -> np.ndarray:
-        """Run a single forward pass and return [N, 5] dets in ORIGINAL image coords."""
-        x, ratio, (dw, dh) = self._preprocess(image)
-        out = self.session.run(self.output_names, {self.input_name: x})[0]
-        if out.ndim == 3:
-            out = out[0]
-        if out.shape[1] < 5:
-            return np.zeros((0, 5), dtype=np.float32)
-        boxes = out[:, :4].astype(np.float32)
-        scores = out[:, 4].astype(np.float32)
-        keep = scores >= conf
-        boxes, scores = boxes[keep], scores[keep]
-        if len(boxes) == 0:
-            return np.zeros((0, 5), dtype=np.float32)
-        boxes[:, [0, 2]] -= dw
-        boxes[:, [1, 3]] -= dh
-        boxes /= ratio
-        oh, ow = image.shape[:2]
-        boxes = self._clip_boxes(boxes, (ow, oh))
-        return np.concatenate([boxes, scores[:, None]], axis=1)
-
-    # ---------- primary pass: soft-NMS + hflip TTA ----------
-    def _primary(self, image: ndarray) -> np.ndarray:
-        d1 = self._raw_dets(image, self.conf_thres)
-        flipped = cv2.flip(image, 1)
-        d2 = self._raw_dets(flipped, self.conf_thres)
-        if len(d2):
-            w = image.shape[1]
-            x1 = w - d2[:, 2]
-            x2 = w - d2[:, 0]
-            d2 = np.stack([x1, d2[:, 1], x2, d2[:, 3], d2[:, 4]], axis=1)
-        all_d = np.concatenate([d1, d2], axis=0) if len(d2) else d1
-        if len(all_d) == 0:
-            return np.zeros((0, 5), dtype=np.float32)
-        # soft-NMS, then hard-NMS
-        keep_idx, scores = self._soft_nms(all_d[:, :4].copy(), all_d[:, 4].copy(), sigma=self.sigma)
-        if len(keep_idx) == 0:
-            return np.zeros((0, 5), dtype=np.float32)
-        merged = np.concatenate([all_d[keep_idx, :4], scores[:, None]], axis=1)
-        keep = self._hard_nms(merged[:, :4], merged[:, 4], self.iou_thres)
-        merged = merged[keep]
-        if len(merged) > self.max_det:
-            merged = merged[np.argsort(-merged[:, 4])[: self.max_det]]
-        return merged
-
-    # ---------- conditional tile pass ----------
-    def _tile_augment(self, image: ndarray, primary: np.ndarray) -> np.ndarray:
-        """Run 2x2 overlapping tiles + hflip, novelty-merge into primary."""
-        oh, ow = image.shape[:2]
-        tw, th = ow // 2, oh // 2
-        ox, oy = int(tw * self.tile_overlap), int(th * self.tile_overlap)
-        tiles = [
-            (0, 0, min(ow, tw + ox), min(oh, th + oy)),
-            (max(0, tw - ox), 0, ow, min(oh, th + oy)),
-            (0, max(0, th - oy), min(ow, tw + ox), oh),
-            (max(0, tw - ox), max(0, th - oy), ow, oh),
-        ]
-        collected: list[np.ndarray] = []
-        for x1, y1, x2, y2 in tiles:
-            crop = image[y1:y2, x1:x2]
-            if crop.size == 0:
-                continue
-            d = self._raw_dets(crop, self.tile_conf)
-            if len(d):
-                d[:, 0] += x1
-                d[:, 1] += y1
-                d[:, 2] += x1
-                d[:, 3] += y1
-                collected.append(d)
-
-        # hflip tile pass (skipped when tile_use_hflip=False — saves 4 ONNX forwards)
-        if self.tile_use_hflip:
-            flipped = cv2.flip(image, 1)
-            for x1, y1, x2, y2 in tiles:
-                fx1 = ow - x2
-                fx2 = ow - x1
-                if fx2 <= fx1:
-                    continue
-                crop = flipped[y1:y2, fx1:fx2]
-                if crop.size == 0:
-                    continue
-                d = self._raw_dets(crop, self.tile_conf)
-                if len(d):
-                    d_un = d.copy()
-                    d_un[:, 0] = (ow - (d[:, 2] + fx1))
-                    d_un[:, 2] = (ow - (d[:, 0] + fx1))
-                    d_un[:, 1] = d[:, 1] + y1
-                    d_un[:, 3] = d[:, 3] + y1
-                    collected.append(d_un)
-
-        if not collected:
-            return primary
-
-        tile_dets = np.concatenate(collected, axis=0)
-        keep = self._hard_nms(tile_dets[:, :4], tile_dets[:, 4], 0.5)
-        tile_dets = tile_dets[keep]
-
-        # Novelty: drop tile boxes that overlap any primary box at IoU >= novelty_iou
-        if len(primary) > 0 and len(tile_dets) > 0:
-            mask = np.ones(len(tile_dets), dtype=bool)
-            for i in range(len(tile_dets)):
-                ious = self._box_iou_one_to_many(tile_dets[i, :4], primary[:, :4])
-                if len(ious) and np.max(ious) >= self.novelty_iou:
-                    mask[i] = False
-            tile_dets = tile_dets[mask]
-
-        if len(tile_dets) == 0:
-            return primary
-
-        # Sanity filter: min/max size, aspect ratio
-        w = tile_dets[:, 2] - tile_dets[:, 0]
-        h = tile_dets[:, 3] - tile_dets[:, 1]
-        area = w * h
-        ar = np.maximum(w / np.maximum(h, 1e-6), h / np.maximum(w, 1e-6))
-        img_area = float(ow * oh)
-        ok = (w >= 6) & (h >= 6) & (area >= 36) & (area <= 0.5 * img_area) & (ar <= 10.0)
-        tile_dets = tile_dets[ok]
-        if len(tile_dets) == 0:
-            return primary
-
-        merged = np.concatenate([primary, tile_dets], axis=0)
-        keep = self._hard_nms(merged[:, :4], merged[:, 4], self.iou_thres)
-        merged = merged[keep]
-        if len(merged) > self.final_max_det:
-            merged = merged[np.argsort(-merged[:, 4])[: self.final_max_det]]
-        return merged
-
-    # ---------- single-image predict ----------
-    def _predict_single(self, image: ndarray) -> list[BoundingBox]:
-        if image is None or not isinstance(image, np.ndarray) or image.ndim != 3:
-            return []
-        if image.shape[0] <= 0 or image.shape[1] <= 0 or image.shape[2] != 3:
-            return []
-        if image.dtype != np.uint8:
-            image = image.astype(np.uint8)
-
-        primary = self._primary(image)
-        if len(primary) < self.sparse_threshold:
-            dets = self._tile_augment(image, primary)
-        else:
-            dets = primary
-
-        results: list[BoundingBox] = []
-        for row in dets:
-            x1, y1, x2, y2, conf = row.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=0,
-                    conf=float(conf),
-                )
-            )
-        return results
-
-    # ---------- chute entrypoint ----------
-    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:
-                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
+# Person detection miner for TurboVision element manak0/Detect-Person.
+# Seeded from the current champion alfred8995/york004 (rev 79eec28…) as a
+# well-proven scoring-aware pipeline. Differentiation is expected to come from:
+#   1. A stronger base detector (YOLO26-s / YOLO11-m vs champion's YOLOv11-nano)
+#   2. Better training data
+#   3. Post-training threshold sweep on our held-out synthetic set
+# Inference contract (inputs/outputs + weights.onnx filename) is locked to the
+# turbovision chute template — do not rename without updating chute_config.yml.
+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"
+        self.class_names = ["person"]
+        print("ORT version:", ort.__version__)
+
+        try:
+            ort.preload_dlls()
+            print("✅ onnxruntime.preload_dlls() success")
+        except Exception as e:
+            print(f"⚠️ preload_dlls failed: {e}")
+
+        print("ORT available providers BEFORE session:", ort.get_available_providers())
+
+        sess_options = ort.SessionOptions()
+        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+
+        try:
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+            )
+            print("✅ Created ORT session with preferred CUDA provider list")
+        except Exception as e:
+            print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}")
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )
+
+        print("ORT session providers:", self.session.get_providers())
+
+        for inp in self.session.get_inputs():
+            print("INPUT:", inp.name, inp.shape, inp.type)
+
+        for out in self.session.get_outputs():
+            print("OUTPUT:", out.name, out.shape, out.type)
+
+        self.input_name = self.session.get_inputs()[0].name
+        self.output_names = [output.name for output in self.session.get_outputs()]
+        self.input_shape = self.session.get_inputs()[0].shape
+
+        self.input_height = self._safe_dim(self.input_shape[2], default=1280)
+        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
+
+        # --- Scoring-aware adaptive confidence ---
+        # total_score = mAP50 * 0.65 + FP_score * 0.35
+        # FP_score = max(0, 1 - n_FP / n_images), typically n_images ≈ 10
+        #
+        # v12s thresholds: swept on SAM3 GT bench (276 imgs). v12s = yolo26-s
+        # student distilled from v12-m teacher (yolo26-m trained on hermestech-
+        # consensus TV labels + 14k aux CCTV/crowd images relabeled by v12-m).
+        # Swept composite 0.8750 vs hermestech 0.8371 (+0.038 lead).
+        self.conf_thres = 0.2149       # Base threshold for candidate generation
+        self.iou_thres = 0.4704         # NMS threshold
+        self.max_det = 150
+
+        # TTA consensus thresholds
+        self.conf_high = 0.7443        # Boxes above this survive without TTA confirmation
+        self.tta_match_iou = 0.4447    # TTA cross-view match IoU
+
+        # Adaptive conf curve: lerp between low/high based on raw detection count
+        self.conf_adapt_low = 0.1467   # Few objects: favor recall
+        self.conf_adapt_high = 0.6997  # Many objects: favor precision
+        self.count_low = 9          # Raw count below this → use conf_adapt_low
+        self.count_high = 46         # Raw count above this → use conf_adapt_high
+
+        self.use_tta = True
+
+        # Box sanity filters
+        self.min_box_area = 14 * 14
+        self.min_w = 8
+        self.min_h = 8
+        self.max_aspect_ratio = 6.5
+        self.max_box_area_ratio = 0.8
+
+        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}")
+
+    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
+        dh = new_h - resized_h
+        dw /= 2.0
+        dh /= 2.0
+
+        left = int(round(dw - 0.1))
+        right = int(round(dw + 0.1))
+        top = int(round(dh - 0.1))
+        bottom = int(round(dh + 0.1))
+
+        padded = cv2.copyMakeBorder(
+            image,
+            top,
+            bottom,
+            left,
+            right,
+            borderType=cv2.BORDER_CONSTANT,
+            value=color,
+        )
+        return padded, ratio, (dw, dh)
+
+    def _preprocess(
+        self, image: ndarray
+    ) -> tuple[np.ndarray, float, tuple[float, float], tuple[int, int]]:
+        orig_h, orig_w = image.shape[:2]
+
+        img, ratio, pad = self._letterbox(
+            image, (self.input_width, self.input_height)
+        )
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        img = img.astype(np.float32) / 255.0
+        img = np.transpose(img, (2, 0, 1))[None, ...]
+        img = np.ascontiguousarray(img, dtype=np.float32)
+
+        return img, ratio, pad, (orig_w, orig_h)
+
+    @staticmethod
+    def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
+        w, h = image_size
+        boxes[:, 0] = np.clip(boxes[:, 0], 0, w - 1)
+        boxes[:, 1] = np.clip(boxes[:, 1], 0, h - 1)
+        boxes[:, 2] = np.clip(boxes[:, 2], 0, w - 1)
+        boxes[:, 3] = np.clip(boxes[:, 3], 0, h - 1)
+        return boxes
+
+    @staticmethod
+    def _xywh_to_xyxy(boxes: np.ndarray) -> np.ndarray:
+        out = np.empty_like(boxes)
+        out[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.0
+        out[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.0
+        out[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.0
+        out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
+        return out
+
+    @staticmethod
+    def _hard_nms(
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        if len(boxes) == 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 = []
+
+        while len(order) > 0:
+            i = 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)
+
+            area_i = np.maximum(0.0, (boxes[i, 2] - boxes[i, 0])) * np.maximum(0.0, (boxes[i, 3] - boxes[i, 1]))
+            area_r = np.maximum(0.0, (boxes[rest, 2] - boxes[rest, 0])) * np.maximum(0.0, (boxes[rest, 3] - boxes[rest, 1]))
+
+            iou = inter / (area_i + area_r - inter + 1e-7)
+            order = rest[iou <= iou_thresh]
+
+        return np.array(keep, dtype=np.intp)
+
+    @staticmethod
+    def _box_iou_one_to_many(box: np.ndarray, boxes: np.ndarray) -> np.ndarray:
+        xx1 = np.maximum(box[0], boxes[:, 0])
+        yy1 = np.maximum(box[1], boxes[:, 1])
+        xx2 = np.minimum(box[2], boxes[:, 2])
+        yy2 = np.minimum(box[3], boxes[:, 3])
+
+        inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+
+        area_a = max(0.0, (box[2] - box[0]) * (box[3] - box[1]))
+        area_b = np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) * np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
+
+        return inter / (area_a + area_b - inter + 1e-7)
+
+    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]:
+        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_w or bh < self.min_h:
+                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),
+            )
+
+        keep = np.array(keep, dtype=np.intp)
+        return boxes[keep], scores[keep], cls_ids[keep]
+
+    def _decode_final_dets(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        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)
+
+        # person only
+        keep = cls_ids == 0
+        boxes = boxes[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+
+        # candidate threshold
+        keep = scores >= self.conf_thres
+        boxes = boxes[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+
+        if len(boxes) == 0:
+            return []
+
+        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)
+        if len(boxes) == 0:
+            return []
+
+        keep_idx = self._hard_nms(boxes, scores, self.iou_thres)
+        keep_idx = keep_idx[: self.max_det]
+
+        boxes = boxes[keep_idx]
+        scores = scores[keep_idx]
+        cls_ids = cls_ids[keep_idx]
+
+        return [
+            BoundingBox(
+                x1=int(math.floor(box[0])),
+                y1=int(math.floor(box[1])),
+                x2=int(math.ceil(box[2])),
+                y2=int(math.ceil(box[3])),
+                cls_id=int(cls_id),
+                conf=float(conf),
+            )
+            for box, conf, cls_id in zip(boxes, scores, cls_ids)
+            if box[2] > box[0] and box[3] > box[1]
+        ]
+
+    def _decode_raw_yolo(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        if preds.ndim != 3:
+            raise ValueError(f"Unexpected raw ONNX output shape: {preds.shape}")
+        if preds.shape[0] != 1:
+            raise ValueError(f"Unexpected batch dimension in raw output: {preds.shape}")
+
+        preds = preds[0]
+
+        # Normalize to [N, C]
+        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 normalized raw output shape: {preds.shape}")
+
+        boxes_xywh = preds[:, :4].astype(np.float32)
+        tail = preds[:, 4:].astype(np.float32)
+
+        # Supports:
+        # [x,y,w,h,score]                 single-class
+        # [x,y,w,h,obj,cls]               YOLO standard single-class
+        # [x,y,w,h,obj,cls1,cls2,...]     multi-class
+        if tail.shape[1] == 1:
+            scores = tail[:, 0]
+            cls_ids = np.zeros(len(scores), dtype=np.int32)
+        elif tail.shape[1] == 2:
+            obj = tail[:, 0]
+            cls_prob = tail[:, 1]
+            scores = obj * cls_prob
+            cls_ids = np.zeros(len(scores), dtype=np.int32)
+        else:
+            obj = tail[:, 0]
+            class_probs = tail[:, 1:]
+            cls_ids = np.argmax(class_probs, axis=1).astype(np.int32)
+            cls_scores = class_probs[np.arange(len(class_probs)), cls_ids]
+            scores = obj * cls_scores
+
+        keep = cls_ids == 0
+        boxes_xywh = boxes_xywh[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+
+        keep = scores >= self.conf_thres
+        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
+        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)
+        if len(boxes) == 0:
+            return []
+
+        keep_idx = self._hard_nms(boxes, scores, self.iou_thres)
+        keep_idx = keep_idx[: self.max_det]
+
+        boxes = boxes[keep_idx]
+        scores = scores[keep_idx]
+        cls_ids = cls_ids[keep_idx]
+
+        return [
+            BoundingBox(
+                x1=int(math.floor(box[0])),
+                y1=int(math.floor(box[1])),
+                x2=int(math.ceil(box[2])),
+                y2=int(math.ceil(box[3])),
+                cls_id=int(cls_id),
+                conf=float(conf),
+            )
+            for box, conf, cls_id in zip(boxes, scores, cls_ids)
+            if box[2] > box[0] and box[3] > box[1]
+        ]
+
+    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_shape = (1, 3, self.input_height, self.input_width)
+        if input_tensor.shape != expected_shape:
+            raise ValueError(
+                f"Bad input tensor shape={input_tensor.shape}, expected={expected_shape}"
+            )
+
+        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
+        det_output = outputs[0]
+        return self._postprocess(det_output, ratio, pad, orig_size)
+
+    def _merge_tta_consensus(
+        self,
+        boxes_orig: list[BoundingBox],
+        boxes_flip: list[BoundingBox],
+    ) -> list[BoundingBox]:
+        """
+        Keep:
+        - any box with conf >= conf_high
+        - low/medium-conf boxes only if confirmed across TTA views
+        Then run final hard NMS.
+        """
+        if not boxes_orig and not boxes_flip:
+            return []
+
+        coords_o = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty((0, 4), dtype=np.float32)
+        scores_o = np.array([b.conf for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty((0,), dtype=np.float32)
+
+        coords_f = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty((0, 4), dtype=np.float32)
+        scores_f = np.array([b.conf for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty((0,), dtype=np.float32)
+
+        accepted_boxes = []
+        accepted_scores = []
+
+        # Original view candidates
+        for i in range(len(coords_o)):
+            score = scores_o[i]
+            if score >= self.conf_high:
+                accepted_boxes.append(coords_o[i])
+                accepted_scores.append(score)
+            elif len(coords_f) > 0:
+                ious = self._box_iou_one_to_many(coords_o[i], coords_f)
+                j = int(np.argmax(ious))
+                if ious[j] >= self.tta_match_iou:
+                    fused_score = max(score, scores_f[j])
+                    accepted_boxes.append(coords_o[i])
+                    accepted_scores.append(fused_score)
+
+        # Flipped-view high-confidence boxes that original missed
+        for i in range(len(coords_f)):
+            score = scores_f[i]
+            if score < self.conf_high:
+                continue
+
+            if len(coords_o) == 0:
+                accepted_boxes.append(coords_f[i])
+                accepted_scores.append(score)
+                continue
+
+            ious = self._box_iou_one_to_many(coords_f[i], coords_o)
+            if np.max(ious) < self.tta_match_iou:
+                accepted_boxes.append(coords_f[i])
+                accepted_scores.append(score)
+
+        if not accepted_boxes:
+            return []
+
+        boxes = np.array(accepted_boxes, dtype=np.float32)
+        scores = np.array(accepted_scores, dtype=np.float32)
+
+        keep = self._hard_nms(boxes, scores, self.iou_thres)
+        keep = keep[: self.max_det]
+
+        out = []
+        for idx in keep:
+            x1, y1, x2, y2 = boxes[idx].tolist()
+            out.append(
+                BoundingBox(
+                    x1=int(math.floor(x1)),
+                    y1=int(math.floor(y1)),
+                    x2=int(math.ceil(x2)),
+                    y2=int(math.ceil(y2)),
+                    cls_id=0,
+                    conf=float(scores[idx]),
+                )
+            )
+        return out
+
+    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        boxes_orig = self._predict_single(image)
+
+        flipped = cv2.flip(image, 1)
+        boxes_flip_raw = 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_raw
+        ]
+
+        return self._merge_tta_consensus(boxes_orig, boxes_flip)
+
+    def _adaptive_conf_threshold(self, n_raw: int) -> float:
+        """
+        Dynamic confidence threshold based on raw detection count.
+
+        total_score = mAP50 * 0.65 + FP_score * 0.35
+        - Few objects → each TP worth ~0.065/n for mAP50 → keep low conf (maximize recall)
+        - Many objects → each TP worth little, FPs dominate → raise conf (minimize FP)
+        """
+        if n_raw <= self.count_low:
+            return self.conf_adapt_low
+        if n_raw >= self.count_high:
+            return self.conf_adapt_high
+        t = (n_raw - self.count_low) / (self.count_high - self.count_low)
+        return self.conf_adapt_low + t * (self.conf_adapt_high - self.conf_adapt_low)
+
+    def _apply_adaptive_filter(self, boxes: list[BoundingBox]) -> list[BoundingBox]:
+        if not boxes:
+            return boxes
+        n_raw = len(boxes)
+        thresh = self._adaptive_conf_threshold(n_raw)
+        return [b for b in boxes if b.conf >= thresh]
+
+    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)
+                boxes = self._apply_adaptive_filter(boxes)
+            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