scorevision: push artifact

miner.py

@@ -1,24 +1,5 @@
-"""
-Detect-vehicle miner for ScoreVision.
-
-Loaded by the TurboVision chute_template from the root of the HF repo.
-Thresholds (imgsz, conf, iou, max_det) are overridable via SN44_* env vars
-so operators can hot-patch without redeploying.
-
-Contract expected by the chute template:
-  * class `Miner(path_hf_repo: Path)`
-  * method `predict_batch(batch_images, offset, n_keypoints) -> list[TVFrameResult]`
-
-Vehicle classes filtered from COCO 80-class: car(2), motorcycle(3), bus(5), truck(7).
-cls_id is REMAPPED to 0 in the output because the ScoreVision validator
-treats class-0 as the single target class per element.
-"""
-
-from __future__ import annotations
-
-import math
-import os
 from pathlib import Path
+import math
 
 import cv2
 import numpy as np
@@ -42,401 +23,384 @@ class TVFrameResult(BaseModel):
     keypoints: list[tuple[int, int]]
 
 
-# ---------------------------------------------------------------------------
-# Tuned hyperparameters (override via env for hot-patching without redeploy)
-# ---------------------------------------------------------------------------
-_DEFAULT_WEIGHTS = "weights.onnx"
-_DEFAULT_IMGSZ = 960
-_DEFAULT_CONF = 0.25
-_DEFAULT_IOU = 0.60
-_DEFAULT_MAX_DET = 300
-
-
-def _env_int(name: str, default: int) -> int:
-    try:
-        return int(os.environ.get(name, default))
-    except (TypeError, ValueError):
-        return default
-
-
-def _env_float(name: str, default: float) -> float:
-    try:
-        return float(os.environ.get(name, default))
-    except (TypeError, ValueError):
-        return default
-
-
-def _letterbox(
-    image: ndarray,
-    new_shape: tuple[int, int],
-    color: tuple[int, int, int] = (114, 114, 114),
-) -> tuple[ndarray, float, tuple[float, float]]:
-    """YOLO-style letterbox preserving aspect ratio, returns (img, ratio, (dw, dh))."""
-    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 _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 _hard_nms(boxes: np.ndarray, scores: np.ndarray, iou_thresh: float) -> np.ndarray:
-    """Pure numpy hard NMS. Avoids torchvision to keep the chute slim."""
-    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: 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)
-        area_i = max(0.0, (boxes[i, 2] - boxes[i, 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)
-
-
-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
+SIZE = 1280
 
 
 class Miner:
-    """Detect-vehicle miner: ONNX Runtime + raw YOLO decode + numpy NMS.
-
-    Same yolo11s_fp16 backbone as the Detect-Person miner; inference path is
-    unchanged. The ONLY difference is class filtering: we keep COCO classes
-    {car=2, motorcycle=3, bus=5, truck=7} and merge them into a single
-    emitted class_id=0 per the validator's single-target convention.
-    """
-
     def __init__(self, path_hf_repo: Path) -> None:
-        # IMPORTANT: element `manak0/Detect-detect-vehicle` declares:
-        #   objects = ["bus", "car", "truck", "motorcycle"]  (in this order)
-        # So the validator maps cls_id 0->"bus", 1->"car", 2->"truck", 3->"motorcycle".
-        # We must emit cls_id matching this order.
-        self.class_names = ["bus", "car", "truck", "motorcycle"]
-        # COCO -> element cls_id:
-        #   bus (5)        -> 0
-        #   car (2)        -> 1
-        #   truck (7)      -> 2
-        #   motorcycle (3) -> 3
-        self.coco_to_element = {5: 0, 2: 1, 7: 2, 3: 3}
-        self.vehicle_coco_ids = tuple(self.coco_to_element.keys())
-
-        weights_name = os.environ.get("SN44_ONNX_WEIGHTS", _DEFAULT_WEIGHTS)
-        weights_path = path_hf_repo / weights_name
-        if not weights_path.is_file():
-            raise FileNotFoundError(
-                f"ONNX weights '{weights_name}' not found in {path_hf_repo}"
-            )
-
+        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("ORT preload_dlls ok")
+            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"ORT preload_dlls skipped: {e}")
-        print("ORT available providers:", ort.get_available_providers())
+            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(weights_path),
+                str(model_path),
                 sess_options=sess_options,
                 providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
             )
-            print("ORT session created with CUDA preferred")
+            print("Created ORT session with preferred CUDA provider list")
         except Exception as e:
-            print(f"ORT CUDA provider failed, falling back to CPU: {e}")
+            print(f"CUDA session creation failed, falling back to CPU: {e}")
             self.session = ort.InferenceSession(
-                str(weights_path),
+                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("ONNX INPUT:", inp.name, inp.shape, inp.type)
+            print("INPUT:", inp.name, inp.shape, inp.type)
         for out in self.session.get_outputs():
-            print("ONNX OUTPUT:", out.name, out.shape, out.type)
+            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()]
-        input_shape = self.session.get_inputs()[0].shape
-
-        h = input_shape[2] if isinstance(input_shape[2], int) and input_shape[2] > 0 else _DEFAULT_IMGSZ
-        w = input_shape[3] if isinstance(input_shape[3], int) and input_shape[3] > 0 else _DEFAULT_IMGSZ
-        self.input_height = _env_int("SN44_IMGSZ", h)
-        self.input_width = _env_int("SN44_IMGSZ", w)
-
-        self.conf_thres = _env_float("SN44_CONF", _DEFAULT_CONF)
-        self.iou_thres = _env_float("SN44_IOU", _DEFAULT_IOU)
-        self.max_det = _env_int("SN44_MAX_DET", _DEFAULT_MAX_DET)
-
-        self.min_w = 4
-        self.min_h = 4
-        self.min_box_area = 16
-        self.max_aspect_ratio = 8.0
-        self.max_box_area_ratio = 0.9
-
-        print(
-            "Vehicle Miner ready: "
-            f"imgsz={self.input_height}x{self.input_width}, "
-            f"conf={self.conf_thres:.3f}, iou={self.iou_thres:.3f}, "
-            f"max_det={self.max_det}, providers={self.session.get_providers()}, "
-            f"coco_ids={self.vehicle_coco_ids}"
-        )
+        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 = 0.23
+        self.iou_thres = 0.66
+        self.sigma = 0.465
+        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 = 0.57
+        self.tile_overlap = 0.20
+        self.novelty_iou = 0.10
+        self.final_max_det = 17
+        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 (
-            "DetectVehicleMiner("
-            f"providers={self.session.get_providers()}, "
-            f"imgsz={self.input_height}x{self.input_width}, "
-            f"conf={self.conf_thres}, iou={self.iou_thres}, "
-            f"coco_ids={self.vehicle_coco_ids})"
+            f"ONNXRuntime(session={type(self.session).__name__}, "
+            f"providers={self.session.get_providers()})"
         )
 
-    def _preprocess(
-        self, image: ndarray
-    ) -> tuple[np.ndarray, float, tuple[float, float], tuple[int, int]]:
-        if image.dtype != np.uint8:
-            image = image.astype(np.uint8)
-        orig_h, orig_w = image.shape[:2]
-        img, ratio, pad = _letterbox(image, (self.input_width, self.input_height))
-        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-        img = img.astype(np.float32) / 255.0
+    @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, ...]
-        img = np.ascontiguousarray(img, dtype=np.float32)
-        return img, ratio, pad, (orig_w, orig_h)
+        return np.ascontiguousarray(img, dtype=np.float32), ratio, pad
 
-    def _filter_sane(
+    @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,
-        orig_size: tuple[int, int],
+        sigma: float,
+        score_thresh: float = 0.01,
     ) -> tuple[np.ndarray, np.ndarray]:
-        if len(boxes) == 0:
-            return boxes, scores
-        orig_w, orig_h = orig_size
-        image_area = float(orig_w * orig_h)
-        keep: list[int] = []
-        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),
+        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])
             )
-        keep_idx = np.array(keep, dtype=np.intp)
-        return boxes[keep_idx], scores[keep_idx]
-
-    def _decode_yolov11(
-        self,
-        preds: np.ndarray,
-        ratio: float,
-        pad: tuple[float, float],
-        orig_size: tuple[int, int],
-    ) -> list[BoundingBox]:
-        """
-        Ultralytics YOLOv8/11 ONNX output is [1, 4+nc, N].
-        For COCO nc=80 → shape [1, 84, N]. No objectness term;
-        class score IS the detection score.
-        """
-        if preds.ndim != 3:
-            return []
-        preds = preds[0]
-        if preds.shape[0] == 4 + len(self._coco_classes()):
-            preds = preds.T
-        elif preds.shape[1] == 4 + len(self._coco_classes()):
-            pass
-        else:
-            if preds.shape[0] < preds.shape[1]:
-                preds = preds.T
-
-        if preds.shape[1] < 5:
-            return []
-
-        boxes_xywh = preds[:, :4].astype(np.float32)
-        class_scores = preds[:, 4:].astype(np.float32)
-
-        # For each detection, determine which COCO vehicle class has the MAX score
-        # and retain that as the class id + score.
-        vehicle_coco_idx = np.array(self.vehicle_coco_ids, dtype=np.intp)
-        vehicle_class_scores = class_scores[:, vehicle_coco_idx]  # (N, 4)
-        best_in_vehicle = vehicle_class_scores.argmax(axis=1)     # (N,) index into vehicle_coco_idx
-        vehicle_scores = vehicle_class_scores.max(axis=1)         # (N,)
-        best_coco_ids = vehicle_coco_idx[best_in_vehicle]         # COCO ids per row
-
-        mask = vehicle_scores >= self.conf_thres
-        if not np.any(mask):
-            return []
-
-        boxes_xywh = boxes_xywh[mask]
-        scores = vehicle_scores[mask]
-        # Map COCO id -> element cls_id (for emission)
-        element_cls_ids = np.array(
-            [self.coco_to_element[int(c)] for c in best_coco_ids[mask]],
-            dtype=np.int32,
-        )
-
-        boxes = _xywh_to_xyxy(boxes_xywh)
-
-        pad_w, pad_h = pad
-        boxes[:, [0, 2]] -= pad_w
-        boxes[:, [1, 3]] -= pad_h
-        boxes /= ratio
-        boxes = _clip_boxes(boxes, orig_size)
+            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]
 
-        boxes, scores, element_cls_ids = self._filter_sane_with_cls(
-            boxes, scores, element_cls_ids, orig_size,
-        )
+    @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 >= 7) & (h >= 7) & (area >= 85) & (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)
 
-        keep = _hard_nms(boxes, scores, self.iou_thres)
-        keep = keep[: self.max_det]
-        boxes = boxes[keep]
-        scores = scores[keep]
-        element_cls_ids = element_cls_ids[keep]
+        primary = self._primary(image)
+        if len(primary) < self.sparse_threshold:
+            dets = self._tile_augment(image, primary)
+        else:
+            dets = primary
 
-        out: list[BoundingBox] = []
-        for box, conf, cls_id in zip(boxes, scores, element_cls_ids):
-            if box[2] <= box[0] or box[3] <= box[1]:
+        results: list[BoundingBox] = []
+        for row in dets:
+            x1, y1, x2, y2, conf = row.tolist()
+            if x2 <= x1 or y2 <= y1:
                 continue
-            out.append(
+            results.append(
                 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),
+                    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 out
-
-    def _filter_sane_with_cls(
-        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: list[int] = []
-        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),
-            )
-        idx = np.array(keep, dtype=np.intp)
-        return boxes[idx], scores[idx], cls_ids[idx]
-
-    @staticmethod
-    def _coco_classes() -> list[str]:
-        return [
-            "person", "bicycle", "car", "motorcycle", "airplane", "bus", "train",
-            "truck", "boat", "traffic light", "fire hydrant", "stop sign",
-            "parking meter", "bench", "bird", "cat", "dog", "horse", "sheep",
-            "cow", "elephant", "bear", "zebra", "giraffe", "backpack", "umbrella",
-            "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard",
-            "sports ball", "kite", "baseball bat", "baseball glove", "skateboard",
-            "surfboard", "tennis racket", "bottle", "wine glass", "cup", "fork",
-            "knife", "spoon", "bowl", "banana", "apple", "sandwich", "orange",
-            "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair",
-            "couch", "potted plant", "bed", "dining table", "toilet", "tv",
-            "laptop", "mouse", "remote", "keyboard", "cell phone", "microwave",
-            "oven", "toaster", "sink", "refrigerator", "book", "clock", "vase",
-            "scissors", "teddy bear", "hair drier", "toothbrush",
-        ]
-
-    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) or image.ndim != 3 or image.shape[2] != 3:
-            raise ValueError(f"Expected HWC RGB/BGR image, got shape={getattr(image, 'shape', None)}")
-
-        input_tensor, ratio, pad, orig_size = self._preprocess(image)
-        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
-        return self._decode_yolov11(outputs[0], ratio, pad, orig_size)
+        return results
 
+    # ---------- chute entrypoint ----------
     def predict_batch(
         self,
         batch_images: list[ndarray],
@@ -444,16 +408,15 @@ class Miner:
         n_keypoints: int,
     ) -> list[TVFrameResult]:
         results: list[TVFrameResult] = []
-        for i, image in enumerate(batch_images):
-            frame_id = offset + i
+        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 {frame_id}: {e}")
+                print(f"Inference failed for frame {offset + frame_number_in_batch}: {e}")
                 boxes = []
             results.append(
                 TVFrameResult(
-                    frame_id=frame_id,
+                    frame_id=offset + frame_number_in_batch,
                     boxes=boxes,
                     keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
                 )