Upload folder using huggingface_hub

chute_config.yml

@@ -0,0 +1,20 @@
+Image:
+  from_base: parachutes/python:3.12
+  run_command:
+    - pip install --upgrade setuptools wheel
+    - pip install 'numpy>=1.23' 'onnxruntime-gpu>=1.16' 'opencv-python>=4.7' 'pillow>=9.5' 'huggingface_hub>=0.19.4' 'pydantic>=2.0' 'pyyaml>=6.0' 'aiohttp>=3.9'
+    - pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
+
+NodeSelector:
+  gpu_count: 1
+  min_vram_gb_per_gpu: 16
+  include:
+    - pro_6000
+
+Chute:
+  timeout_seconds: 900
+  concurrency: 4
+  max_instances: 5
+  scaling_threshold: 0.5
+  shutdown_after_seconds: 288000
+  tee: true

class_names.txt

@@ -0,0 +1,3 @@
+bottle
+can
+cup

miner.py

@@ -0,0 +1,869 @@
+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 = ['cup', 'bottle', 'can']
+        self._cls_cup = self.class_names.index("cup")
+        self._cls_bottle = self.class_names.index("bottle")
+        self._cls_can = self.class_names.index("can")
+        model_class_order = ["bottle", "can", "cup"]
+        self.cls_remap = np.array(
+            [self.class_names.index(n) for n in model_class_order], 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())
+
+        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
+
+        # Your export is fixed-size 1280, but we still read actual ONNX input shape first.
+        self.input_height = self._safe_dim(self.input_shape[2], default=1280)
+        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
+
+        # Tuned for validator scoring: reduce FP (FALSE_POSITIVE pillar),
+        # preserve recall (MAP50, RECALL), improve precision.
+        self.conf_thres = 0.32  # Higher = fewer FP, slightly lower recall
+        self.iou_thres = 0.5   # Lower = suppress duplicate detections (FP)
+        self.cross_iou_thresh = 0.6
+        self.max_det = 150     # Cap detections per image
+        self.use_tta = True
+
+        # Box sanity: filter tiny/spurious detections (common FP source)
+        self.min_box_area = 100  # ~144 px²
+        self.min_side = 6
+        self.max_aspect_ratio = 8.0
+
+        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]]:
+        """
+        Resize with unchanged aspect ratio and pad to target shape.
+        Returns:
+            padded_image,
+            ratio,
+            (pad_w, pad_h)  # half-padding
+        """
+        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]]:
+        """
+        Preprocess for fixed-size ONNX export:
+        - enhance image quality (CLAHE, denoise, sharpen)
+        - letterbox to model input size
+        - BGR -> RGB
+        - normalize to [0,1]
+        - HWC -> NCHW float32
+        """
+        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
+
+    def _soft_nms(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        sigma: float = 0.5,
+        score_thresh: float = 0.01,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """
+        Soft-NMS: Gaussian decay of overlapping scores instead of hard removal.
+        Processing order prefers **larger** boxes first (then score), so duplicate
+        detections on one object tend to keep the larger box.
+        Returns (kept_original_indices, updated_scores).
+        """
+        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)
+        areas = (
+            np.maximum(0.0, boxes[:, 2] - boxes[:, 0])
+            * np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
+        ).astype(np.float32)
+        order = np.arange(N)
+
+        for i in range(N):
+            max_pos = i + int(np.lexsort((-scores[i:], -areas[i:]))[-1])
+            boxes[[i, max_pos]] = boxes[[max_pos, i]]
+            scores[[i, max_pos]] = scores[[max_pos, i]]
+            order[[i, max_pos]] = order[[max_pos, i]]
+            areas[[i, max_pos]] = areas[[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 = max(0.0, 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 _hard_nms(
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        """
+        Hard NMS: keep one box per overlapping cluster.
+        When two boxes cover the same object, keep the **larger** box (area),
+        breaking ties with higher score.
+        """
+        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)
+        areas = np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) * np.maximum(
+            0.0, boxes[:, 3] - boxes[:, 1]
+        )
+        order = np.lexsort((-scores, -areas))
+        keep: list[int] = []
+        suppressed = np.zeros(N, dtype=bool)
+        for i in range(N):
+            idx = order[i]
+            if suppressed[idx]:
+                continue
+            keep.append(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)
+
+    def _per_class_hard_nms(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        """Hard NMS applied independently per class."""
+        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 _per_class_soft_nms(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        sigma: float = 0.5,
+        score_thresh: float = 0.01,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """Soft NMS applied independently per class."""
+        if len(boxes) == 0:
+            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
+        all_keep: list[int] = []
+        all_scores: list[float] = []
+        for c in np.unique(cls_ids):
+            mask = cls_ids == c
+            indices = np.where(mask)[0]
+            keep, updated = self._soft_nms(boxes[mask], scores[mask], sigma, score_thresh)
+            for k, s in zip(keep, updated):
+                all_keep.append(int(indices[k]))
+                all_scores.append(float(s))
+        if not all_keep:
+            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
+        return np.array(all_keep, dtype=np.intp), np.array(all_scores, dtype=np.float32)
+
+    @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]:
+        """Suppress high-overlap duplicates regardless of class."""
+        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]
+        )
+        # Match dataset-prep behavior: 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 _iou_xyxy(a: np.ndarray, b: np.ndarray) -> float:
+        """Intersection-over-union for two xyxy boxes (float arrays length 4)."""
+        ax1, ay1, ax2, ay2 = float(a[0]), float(a[1]), float(a[2]), float(a[3])
+        bx1, by1, bx2, by2 = float(b[0]), float(b[1]), float(b[2]), float(b[3])
+        ix1 = max(ax1, bx1)
+        iy1 = max(ay1, by1)
+        ix2 = min(ax2, bx2)
+        iy2 = min(ay2, by2)
+        iw = max(0.0, ix2 - ix1)
+        ih = max(0.0, iy2 - iy1)
+        inter = iw * ih
+        area_a = max(0.0, ax2 - ax1) * max(0.0, ay2 - ay1)
+        area_b = max(0.0, bx2 - bx1) * max(0.0, by2 - by1)
+        union = area_a + area_b - inter + 1e-7
+        return inter / union
+
+    def _apply_cross_class_precedence(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        iou_thresh: float | None = None,
+    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        When one object is detected as multiple classes (high IoU overlap):
+        - bottle wins over cup and can (drop overlapping cup / can)
+        - can wins over cup (drop overlapping cup when no bottle conflict)
+        """
+        thr = self.cross_iou_thresh if iou_thresh is None else iou_thresh
+        if len(boxes) == 0:
+            return boxes, scores, cls_ids
+
+        bottle_boxes = boxes[cls_ids == self._cls_bottle]
+        can_boxes = boxes[cls_ids == self._cls_can]
+        cup_mask = cls_ids == self._cls_cup
+        can_mask = cls_ids == self._cls_can
+
+        keep_row = np.ones(len(boxes), dtype=bool)
+
+        # Can loses to bottle
+        if len(bottle_boxes) > 0 and can_mask.any():
+            for i in np.where(can_mask)[0]:
+                bi = boxes[i]
+                for bb in bottle_boxes:
+                    if self._iou_xyxy(bi, bb) >= thr:
+                        keep_row[i] = False
+                        break
+
+        # Cup loses to bottle or can
+        if cup_mask.any():
+            for i in np.where(cup_mask)[0]:
+                if not keep_row[i]:
+                    continue
+                bi = boxes[i]
+                if len(bottle_boxes) > 0:
+                    for bb in bottle_boxes:
+                        if self._iou_xyxy(bi, bb) >= thr:
+                            keep_row[i] = False
+                            break
+                if keep_row[i] and len(can_boxes) > 0:
+                    for cb in can_boxes:
+                        if self._iou_xyxy(bi, cb) >= thr:
+                            keep_row[i] = False
+                            break
+
+        if keep_row.all():
+            return boxes, scores, cls_ids
+        k = np.where(keep_row)[0]
+        return boxes[k], scores[k], cls_ids[k]
+
+    def _apply_cross_class_precedence_list(
+        self, boxes: list[BoundingBox]
+    ) -> list[BoundingBox]:
+        """Same precedence as _apply_cross_class_precedence for post-TTA lists."""
+        if len(boxes) < 2:
+            return boxes
+        thr = self.cross_iou_thresh
+        bottles = [b for b in boxes if b.cls_id == self._cls_bottle]
+        cans = [b for b in boxes if b.cls_id == self._cls_can]
+
+        def overlaps_any(ba: np.ndarray, others: list[BoundingBox]) -> bool:
+            for o in others:
+                oa = np.array([o.x1, o.y1, o.x2, o.y2], dtype=np.float32)
+                if self._iou_xyxy(ba, oa) >= thr:
+                    return True
+            return False
+
+        out: list[BoundingBox] = []
+        for b in boxes:
+            ba = np.array([b.x1, b.y1, b.x2, b.y2], dtype=np.float32)
+            if b.cls_id == self._cls_can:
+                if bottles and overlaps_any(ba, bottles):
+                    continue
+            elif b.cls_id == self._cls_cup:
+                if bottles and overlaps_any(ba, bottles):
+                    continue
+                if cans and overlaps_any(ba, cans):
+                    continue
+            out.append(b)
+        return out
+
+    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]:
+        """Filter out tiny, degenerate, or implausible boxes (common 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]
+
+    @staticmethod
+    def _max_score_per_cluster(
+        coords: np.ndarray,
+        scores: np.ndarray,
+        keep_indices: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        """
+        For each kept box, return the max original score among itself and any
+        box that overlaps it with IOU >= iou_thresh (so TTA cluster keeps best conf).
+        """
+        n_keep = len(keep_indices)
+        if n_keep == 0:
+            return np.array([], dtype=np.float32)
+        out = np.empty(n_keep, dtype=np.float32)
+        coords = np.asarray(coords, dtype=np.float32)
+        scores = np.asarray(scores, 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_final_dets(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+        apply_optional_dedup: bool = False,
+    ) -> list[BoundingBox]:
+        """
+        Primary path:
+        expected output rows like [x1, y1, x2, y2, conf, cls_id]
+        in letterboxed input coordinates.
+        """
+        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 = 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
+
+        # reverse letterbox
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, (orig_w, orig_h))
+
+        # Box sanity filter (reduces FP)
+        boxes, scores, cls_ids = self._filter_sane_boxes(
+            boxes, scores, cls_ids, orig_size
+        )
+        if len(boxes) == 0:
+            return []
+
+        # Per-class NMS to remove duplicates without suppressing across classes
+        if len(boxes) > 1:
+            if apply_optional_dedup:
+                keep_idx, scores = self._per_class_soft_nms(boxes, scores, cls_ids)
+                boxes = boxes[keep_idx]
+                cls_ids = cls_ids[keep_idx]
+            else:
+                keep_idx = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
+                keep_idx = keep_idx[: self.max_det]
+                boxes = boxes[keep_idx]
+                scores = scores[keep_idx]
+                cls_ids = cls_ids[keep_idx]
+            boxes, scores, cls_ids = self._cross_class_dedup(
+                boxes, scores, cls_ids, self.cross_iou_thresh
+            )
+
+        if len(boxes) > 0:
+            boxes, scores, cls_ids = self._apply_cross_class_precedence(
+                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 _decode_raw_yolo(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """
+        Fallback path for raw YOLO predictions.
+        Supports common layouts:
+        - [1, C, N]
+        - [1, N, C]
+        """
+        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)
+        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 = 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)
+
+        keep_idx = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
+        keep_idx = keep_idx[: self.max_det]
+        boxes = boxes[keep_idx]
+        scores = scores[keep_idx]
+        cls_ids = cls_ids[keep_idx]
+        boxes, scores, cls_ids = self._cross_class_dedup(
+            boxes, scores, cls_ids, self.cross_iou_thresh
+        )
+        boxes, scores, cls_ids = self._apply_cross_class_precedence(
+            boxes, scores, cls_ids
+        )
+
+        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_w, orig_h)
+        )
+        if len(boxes) == 0:
+            return []
+
+        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 _postprocess(
+        self,
+        output: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """
+        Prefer final detections first.
+        Fallback to raw decode only if needed.
+        """
+        # final detections: [N,6]
+        if output.ndim == 2 and output.shape[1] >= 6:
+            return self._decode_final_dets(output, ratio, pad, orig_size)
+
+        # final detections: [1,N,6]
+        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 6:
+            return self._decode_final_dets(output, ratio, pad, orig_size)
+
+        # fallback raw decode
+        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 _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        """
+        Horizontal-flip TTA: merge original + flipped via hard NMS.
+        Boost confidence for consensus detections (both views agree) to improve
+        mAP: validator sorts by confidence, so higher conf for TP helps PR curve.
+        """
+        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 len(all_boxes) == 0:
+            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)
+
+        hard_keep = self._per_class_hard_nms(coords, scores, cls_ids, self.iou_thres)
+        if len(hard_keep) == 0:
+            return []
+
+        hard_keep = hard_keep[: self.max_det]
+
+        # Boost confidence when both views agree (overlapping detections)
+        boosted = self._max_score_per_cluster(
+            coords, scores, hard_keep, self.iou_thres
+        )
+
+        return self._apply_cross_class_precedence_list(
+            [
+                BoundingBox(
+                    x1=all_boxes[i].x1,
+                    y1=all_boxes[i].y1,
+                    x2=all_boxes[i].x2,
+                    y2=all_boxes[i].y2,
+                    cls_id=all_boxes[i].cls_id,
+                    conf=float(boosted[j]),
+                )
+                for j, i in enumerate(hard_keep)
+            ]
+        )
+
+    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)
+            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

weights.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c0831618841953e228d3e19a3ce1726a7ea2a8650c0bda1600eea1c4ea94da8
+size 19407230