Update miner.py

miner.py

@@ -1,647 +1,565 @@
 from pathlib import Path
+import math
 
+import cv2
+import numpy as np
 import onnxruntime as ort
 from numpy import ndarray
 from pydantic import BaseModel
-import numpy as np
-import cv2
-
-
-MODEL_CLASS_NAMES = ["balaclava", "bat", "glove", "graffiti", "hoodie", "spray paint"]
-CLASS_CONF_LIST = [0.44, 0.25, 0.45, 0.15, 0.45, 0.45]
-BONUS_LIST = [0.23, 0.3, 0.45, 0.44, 0.45, 0.41]
-
-SUBNET_CLASS_NAMES = ["balaclava", "hoodie", "glove", "bat", "spray paint", "graffiti"]
-
-# =========================
-# ONNX YOLO MODEL
-# =========================
-
-
-class ONNXYOLO:
-    def __init__(
-        self,
-        weights_path: str,
-        imgsz: int = 640,
-        providers: list[str] | None = None,
-    ):
-        self.weights_path = weights_path
-        self.imgsz = imgsz
 
-        if providers is None:
-            providers = ["CPUExecutionProvider"]
 
-        session_options = ort.SessionOptions()
-        session_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-        session_options.intra_op_num_threads = 8
-        session_options.inter_op_num_threads = 1
-
-        self.session = ort.InferenceSession(
-            weights_path,
-            sess_options=session_options,
-            providers=providers,
-        )
+class BoundingBox(BaseModel):
+    x1: int
+    y1: int
+    x2: int
+    y2: int
+    cls_id: int
+    conf: float
 
-        self.input_name = self.session.get_inputs()[0].name
-        self.output_names = [output.name for output in self.session.get_outputs()]
 
-        print("ONNX providers:", self.session.get_providers())
-        print("ONNX input:", self.session.get_inputs()[0].name, self.session.get_inputs()[0].shape)
-        print("ONNX outputs:")
-        for output in self.session.get_outputs():
-            print(" ", output.name, output.shape, output.type)
-
-    def letterbox(
-        self,
-        image: np.ndarray,
-        new_shape: int = 640,
-        color: tuple[int, int, int] = (114, 114, 114),
-    ):
-        """
-        Same idea as Ultralytics letterbox:
-        resize image while preserving aspect ratio and pad to square.
-        """
-        original_h, original_w = image.shape[:2]
-
-        scale = min(new_shape / original_h, new_shape / original_w)
-
-        resized_w = int(round(original_w * scale))
-        resized_h = int(round(original_h * scale))
-
-        pad_w = new_shape - resized_w
-        pad_h = new_shape - resized_h
-
-        pad_left = pad_w // 2
-        pad_right = pad_w - pad_left
-        pad_top = pad_h // 2
-        pad_bottom = pad_h - pad_top
-
-        resized = cv2.resize(image, (resized_w, resized_h), interpolation=cv2.INTER_LINEAR)
-
-        padded = cv2.copyMakeBorder(
-            resized,
-            pad_top,
-            pad_bottom,
-            pad_left,
-            pad_right,
-            cv2.BORDER_CONSTANT,
-            value=color,
-        )
+class TVFrameResult(BaseModel):
+    frame_id: int
+    boxes: list[BoundingBox]
+    keypoints: list[tuple[int, int]]
 
-        return padded, scale, pad_left, pad_top
 
-    def preprocess(self, image: np.ndarray):
-        """
-        Input image is OpenCV BGR uint8.
-        Output is float32 NCHW RGB normalized 0..1.
-        """
-        padded, scale, pad_left, pad_top = self.letterbox(
-            image=image,
-            new_shape=self.imgsz,
-        )
+class Miner:
+    """ONNX Runtime miner. Hard per-class NMS + cross-class dedup + flip TTA."""
+
+    class_names = ["balaclava", "hoodie", "glove", "bat", "spray paint", "graffiti"]
+
+    # FALLBACK order the model emits classes in -- remapped to `class_names`
+    # index by `self.cls_remap` (built in __init__). The authoritative order is
+    # read from the ONNX `names` metadata that Ultralytics embeds at export time
+    # (ships inside weights.onnx), so a retrained model with a different class
+    # order is remapped correctly without code changes. This static list is used
+    # only when that metadata is missing or unparsable.
+    model_class_names = ["balaclava", "bat", "glove", "graffiti", "hoodie", "spray paint"]
+
+    input_size = 1280
+
+    # Test-time augmentation (horizontal-flip ensemble) runs a SECOND forward
+    # pass per frame and roughly DOUBLES latency. This 640 model is built for a
+    # <100 ms single-pass budget, so TTA is OFF by default. Turn it on only with
+    # latency headroom — and note the per-class thresholds below should be
+    # re-swept for whichever mode you deploy, since flipping TTA shifts scores.
+    use_tta = False
+
+    iou_thres = 0.3
+    cross_iou_thresh = 0.8
+    max_det = 150
+    
+    _conf_thres_array = np.array(
+        [0.38, 0.58, 0.22, 0.25, 0.23, 0.22], dtype=np.float32,
+    )
+    _bonus_array = np.array(
+        [0.18, 0.25, 0.12, 0.09, 0.07, 0.10], dtype=np.float32,
+    )
 
-        rgb = cv2.cvtColor(padded, cv2.COLOR_BGR2RGB)
-
-        x = rgb.astype(np.float32) / 255.0
-        x = np.transpose(x, (2, 0, 1))
-        x = np.expand_dims(x, axis=0)
-        x = np.ascontiguousarray(x, dtype=np.float32)
-
-        return x, scale, pad_left, pad_top
-
-    def postprocess_raw_yolo(
-        self,
-        output: np.ndarray,
-        original_shape: tuple[int, int],
-        scale: float,
-        pad_left: int,
-        pad_top: int,
-        conf_threshold: float,
-        iou_threshold: float = 0.7,
-    ) -> list[dict]:
-        """
-        Supports common Ultralytics YOLO ONNX raw output:
-
-        Case A:
-            [1, 4 + nc, N]
-            example: [1, 10, 8400] for 6 classes
-
-        Case B:
-            [1, N, 4 + nc]
-
-        Case C with NMS:
-            [1, N, 6] or [N, 6]
-            [x1, y1, x2, y2, conf, cls]
-        """
-        if isinstance(output, list):
-            output = output[0]
-
-        pred = np.asarray(output)
-
-        # Remove batch dim.
-        if pred.ndim == 3 and pred.shape[0] == 1:
-            pred = pred[0]
-
-        # NMS-exported output: [N, 6]
-        if pred.ndim == 2 and pred.shape[1] == 6:
-            return self.parse_nms_output(
-                pred=pred,
-                original_shape=original_shape,
-                scale=scale,
-                pad_left=pad_left,
-                pad_top=pad_top,
-                conf_threshold=conf_threshold,
+    def __init__(self, path_hf_repo: Path) -> None:
+        model_path = path_hf_repo / "weights-all.onnx"
+        print("ORT version:", ort.__version__)
+
+        try:
+            ort.preload_dlls()
+            print("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
+        # Pin threads for the 2vCPU/4GB public-track latency gate (p95 <= 100ms).
+        sess_options.intra_op_num_threads = 2
+        sess_options.inter_op_num_threads = 1
+        sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+
+        try:
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )            
+        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"],
             )
 
-        # Raw output [4 + nc, N] -> [N, 4 + nc]
-        if pred.ndim == 2 and pred.shape[0] == 4 + len(MODEL_CLASS_NAMES):
-            pred = pred.T
-
-        # Raw output [N, 4 + nc]
-        if pred.ndim != 2:
-            raise ValueError(f"Unsupported ONNX output shape: {pred.shape}")
-
-        if pred.shape[1] == 4 + len(MODEL_CLASS_NAMES):
-            boxes_xywh = pred[:, :4]
-            class_scores = pred[:, 4:]
-        elif pred.shape[1] == 5 + len(MODEL_CLASS_NAMES):
-            # YOLOv5-style: x,y,w,h,obj,cls...
-            boxes_xywh = pred[:, :4]
-            objectness = pred[:, 4:5]
-            class_scores = pred[:, 5:] * objectness
+        print("ORT session providers:", self.session.get_providers())
+
+        # Build cls_remap: for each model-emit index i,
+        #   cls_remap[i] = self.class_names.index(model_class_order[i])
+        # i.e. convert a model-side class id into the output class id that
+        # downstream code (BoundingBox.cls_id, the per-class threshold/bonus
+        # arrays) expects. The model-side order comes from the ONNX metadata
+        # when available, else falls back to the static model_class_names.
+        model_class_order = self._read_model_class_order()
+        if model_class_order is None:
+            model_class_order = list(self.model_class_names)
+            print(f"cls order: no usable ONNX metadata, FALLBACK {model_class_order}")
         else:
-            raise ValueError(
-                f"Unsupported prediction shape {pred.shape}. "
-                f"Expected second dimension {4 + len(MODEL_CLASS_NAMES)} or {5 + len(MODEL_CLASS_NAMES)}."
-            )
-
-        class_ids = np.argmax(class_scores, axis=1)
-        confidences = class_scores[np.arange(class_scores.shape[0]), class_ids]
-
-        keep = confidences >= conf_threshold
-
-        boxes_xywh = boxes_xywh[keep]
-        confidences = confidences[keep]
-        class_ids = class_ids[keep]
-
-        detections = []
-
-        original_h, original_w = original_shape
-
-        for xywh, confidence, class_id in zip(boxes_xywh, confidences, class_ids):
-            class_id = int(class_id)
-            confidence = float(confidence)
+            print(f"cls order: from ONNX metadata {model_class_order}")
+        self.cls_remap = np.array(
+            [self.class_names.index(n) for n in model_class_order],
+            dtype=np.int32,
+        )
 
-            if class_id < 0 or class_id >= len(CLASS_CONF_LIST):
-                continue
+        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)
 
-            if confidence < CLASS_CONF_LIST[class_id]:
-                continue
-
-            cx, cy, w, h = xywh
-
-            x1 = cx - w / 2.0
-            y1 = cy - h / 2.0
-            x2 = cx + w / 2.0
-            y2 = cy + h / 2.0
-
-            # Undo letterbox padding and scale.
-            x1 = (x1 - pad_left) / scale
-            x2 = (x2 - pad_left) / scale
-            y1 = (y1 - pad_top) / scale
-            y2 = (y2 - pad_top) / scale
-
-            x1 = np.clip(x1, 0, original_w - 1)
-            x2 = np.clip(x2, 0, original_w - 1)
-            y1 = np.clip(y1, 0, original_h - 1)
-            y2 = np.clip(y2, 0, original_h - 1)
-
-            detections.append(
-                {
-                    "x1": float(x1),
-                    "y1": float(y1),
-                    "x2": float(x2),
-                    "y2": float(y2),
-                    "cls_id": class_id,
-                    "conf": confidence,
-                }
+        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=self.input_size)
+        self.input_width = self._safe_dim(self.input_shape[3], default=self.input_size)
+
+        print(f"ONNX model loaded from: {model_path}")
+        print(f"ONNX providers: {self.session.get_providers()}")
+        print(f"ONNX input: name={self.input_name}, shape={self.input_shape}")
+        print(f"ONNX input size: {self.input_width}x{self.input_height}, use_tta={self.use_tta}")
+        print("per-class conf: " + ", ".join(
+            f"{n}={t:.3f}" for n, t in zip(self.class_names,
+                                           self._conf_thres_array.tolist())))
+
+        self._warmup()
+
+    def _warmup(self, iters: int = 3) -> None:
+        try:
+            dummy = np.zeros((720, 1280, 3), dtype=np.uint8)
+            for _ in range(max(1, iters)):
+                self.predict_batch(batch_images=[dummy], offset=0, n_keypoints=0)
+            print(f"warmup: {iters} dummy predict_batch call(s) done")
+        except Exception as e:
+            print(f"warmup skipped: {e}")
+
+    def _read_model_class_order(self) -> "list[str] | None":
+        """Read the model's class order from Ultralytics ONNX metadata.
+        Returns the class names ordered by model-emit index, or None when the
+        metadata is missing/unparsable or doesn't match `class_names` as a set
+        (in which case the static model_class_names fallback is used)."""
+        try:
+            import ast
+
+            meta = self.session.get_modelmeta().custom_metadata_map
+            names = ast.literal_eval(meta["names"])  # e.g. {0: 'balaclava', ...}
+            if isinstance(names, dict):
+                order = [str(names[i]) for i in sorted(names)]
+            else:
+                order = [str(n) for n in names]
+        except Exception as e:
+            print(f"cls order: could not read ONNX names metadata ({e})")
+            return None
+        if sorted(order) != sorted(self.class_names):
+            print(
+                f"cls order: ONNX names {order} do not match expected classes "
+                f"{self.class_names}; ignoring metadata"
             )
+            return None
+        return order
 
-        detections = nms_same_class(
-            detections=detections,
-            iou_threshold=iou_threshold,
+    def __repr__(self) -> str:
+        return (
+            f"ONNXRuntime(session={type(self.session).__name__}, "
+            f"providers={self.session.get_providers()})"
         )
 
-        return detections
-
-    def parse_nms_output(
-        self,
-        pred: np.ndarray,
-        original_shape: tuple[int, int],
-        scale: float,
-        pad_left: int,
-        pad_top: int,
-        conf_threshold: float,
-    ) -> list[dict]:
-        detections = []
-        original_h, original_w = original_shape
-
-        for row in pred:
-            x1, y1, x2, y2, confidence, class_id = row.tolist()
-
-            class_id = int(class_id)
-            confidence = float(confidence)
-
-            if confidence < conf_threshold:
+    @staticmethod
+    def _safe_dim(value, default: int) -> int:
+        return value if isinstance(value, int) and value > 0 else default
+
+    def _letterbox(self, image: ndarray, new_shape: tuple[int, int],
+                   color=(114, 114, 114)
+                   ) -> tuple[ndarray, float, tuple[float, float]]:
+        h, w = image.shape[:2]
+        new_w, new_h = new_shape
+        ratio = min(new_w / w, new_h / h)
+        resized_w = int(round(w * ratio))
+        resized_h = int(round(h * ratio))
+        if (resized_w, resized_h) != (w, h):
+            interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR
+            image = cv2.resize(image, (resized_w, resized_h), interpolation=interp)
+        dw = (new_w - resized_w) / 2.0
+        dh = (new_h - resized_h) / 2.0
+        left = int(round(dw - 0.1))
+        right = int(round(dw + 0.1))
+        top = int(round(dh - 0.1))
+        bottom = int(round(dh + 0.1))
+        padded = cv2.copyMakeBorder(image, top, bottom, left, right,
+                                    borderType=cv2.BORDER_CONSTANT, value=color)
+        return padded, ratio, (dw, dh)
+
+    def _preprocess(self, image: ndarray
+                    ) -> tuple[np.ndarray, float, tuple[float, float],
+                               tuple[int, int]]:
+        orig_h, orig_w = image.shape[:2]
+        img, ratio, pad = self._letterbox(image, (self.input_width, self.input_height))
+        # Fused scale(1/255) + BGR->RGB swap + HWC->NCHW + contiguous float32 in
+        # one optimized OpenCV call (bit-identical to the cvtColor + astype/255 +
+        # transpose chain, but ~half the preprocess time).
+        blob = cv2.dnn.blobFromImage(img, scalefactor=1.0 / 255.0, swapRB=True)
+        return blob, ratio, pad, (orig_w, orig_h)
+
+    @staticmethod
+    def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
+        w, h = image_size
+        boxes[:, 0] = np.clip(boxes[:, 0], 0, w - 1)
+        boxes[:, 1] = np.clip(boxes[:, 1], 0, h - 1)
+        boxes[:, 2] = np.clip(boxes[:, 2], 0, w - 1)
+        boxes[:, 3] = np.clip(boxes[:, 3], 0, h - 1)
+        return boxes
+
+    @staticmethod
+    def _xywh_to_xyxy(boxes: np.ndarray) -> np.ndarray:
+        out = np.empty_like(boxes)
+        out[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.0
+        out[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.0
+        out[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.0
+        out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
+        return out
+
+    @staticmethod
+    def _hard_nms(boxes: np.ndarray, scores: np.ndarray,
+                  iou_thresh: float) -> np.ndarray:
+        n = len(boxes)
+        if n == 0:
+            return np.array([], dtype=np.intp)
+        order = np.argsort(-scores)
+        keep: list[int] = []
+        while len(order) > 0:
+            i = int(order[0])
+            keep.append(i)
+            if len(order) == 1:
+                break
+            rest = order[1:]
+            xx1 = np.maximum(boxes[i, 0], boxes[rest, 0])
+            yy1 = np.maximum(boxes[i, 1], boxes[rest, 1])
+            xx2 = np.minimum(boxes[i, 2], boxes[rest, 2])
+            yy2 = np.minimum(boxes[i, 3], boxes[rest, 3])
+            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            a_i = (max(0.0, boxes[i, 2] - boxes[i, 0]) *
+                   max(0.0, boxes[i, 3] - boxes[i, 1]))
+            a_r = (np.maximum(0.0, boxes[rest, 2] - boxes[rest, 0]) *
+                   np.maximum(0.0, boxes[rest, 3] - boxes[rest, 1]))
+            iou = inter / (a_i + a_r - inter + 1e-7)
+            order = rest[iou <= iou_thresh]
+        return np.array(keep, dtype=np.intp)
+
+    def _per_class_hard_nms(self, boxes: np.ndarray, scores: np.ndarray,
+                            cls_ids: np.ndarray, iou_thresh: float
+                            ) -> np.ndarray:
+        if len(boxes) == 0:
+            return np.array([], dtype=np.intp)
+        all_keep: list[int] = []
+        for c in np.unique(cls_ids):
+            mask = cls_ids == c
+            indices = np.where(mask)[0]
+            keep = self._hard_nms(boxes[mask], scores[mask], iou_thresh)
+            all_keep.extend(indices[keep].tolist())
+        all_keep.sort()
+        return np.array(all_keep, dtype=np.intp)
+
+    def _cross_class_dedup_op(self, boxes: np.ndarray, scores: np.ndarray,
+                              cls_ids: np.ndarray, iou_thresh: float
+                              ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        n = len(boxes)
+        if n <= 1:
+            return boxes, scores, cls_ids
+        boxes = np.asarray(boxes, dtype=np.float32)
+        scores = np.asarray(scores, dtype=np.float32)
+        cls_ids = np.asarray(cls_ids, dtype=np.int32)
+        areas = (np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) *
+                 np.maximum(0.0, boxes[:, 3] - boxes[:, 1]))
+        margins = scores - self._conf_thres_array[cls_ids]
+        order = np.lexsort((-areas, -margins))
+        suppressed = np.zeros(n, dtype=bool)
+        keep: list[int] = []
+        for i in order:
+            if suppressed[i]:
                 continue
-
-            if class_id < 0 or class_id >= len(CLASS_CONF_LIST):
+            keep.append(int(i))
+            bi = boxes[i]
+            xx1 = np.maximum(bi[0], boxes[:, 0])
+            yy1 = np.maximum(bi[1], boxes[:, 1])
+            xx2 = np.minimum(bi[2], boxes[:, 2])
+            yy2 = np.minimum(bi[3], boxes[:, 3])
+            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            a_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
+            iou = inter / (a_i + areas - inter + 1e-7)
+            dup = iou > iou_thresh
+            dup[i] = False
+            suppressed |= dup
+        keep_idx = np.array(keep, dtype=np.intp)
+        return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx]
+
+    @staticmethod
+    def _max_score_per_cluster(post_boxes: np.ndarray,
+                               post_cls: np.ndarray,
+                               full_boxes: np.ndarray,
+                               full_scores: np.ndarray,
+                               full_cls: np.ndarray,
+                               iou_thresh: float) -> np.ndarray:
+        n = len(post_boxes)
+        if n == 0:
+            return np.empty(0, dtype=np.float32)
+        full_areas = (np.maximum(0.0, full_boxes[:, 2] - full_boxes[:, 0]) *
+                      np.maximum(0.0, full_boxes[:, 3] - full_boxes[:, 1]))
+        out = np.empty(n, dtype=np.float32)
+        for i in range(n):
+            bi = post_boxes[i]
+            xx1 = np.maximum(bi[0], full_boxes[:, 0])
+            yy1 = np.maximum(bi[1], full_boxes[:, 1])
+            xx2 = np.minimum(bi[2], full_boxes[:, 2])
+            yy2 = np.minimum(bi[3], full_boxes[:, 3])
+            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            a_i = max(0.0, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
+            iou = inter / (a_i + full_areas - inter + 1e-7)
+            cluster = (iou >= iou_thresh) & (full_cls == post_cls[i])
+            out[i] = float(np.max(full_scores[cluster])) if np.any(cluster) else 0.0
+        return out
+
+    def _conf_filter_mask(self, scores: np.ndarray,
+                          cls_ids: np.ndarray) -> np.ndarray:
+        """Boolean keep-mask: score >= per-class threshold, with a per-class
+        rescue — if a class has zero boxes passing, admit its top-1 candidate
+        when its score >= (per-class threshold - per-class bonus)."""
+        if len(scores) == 0:
+            return np.zeros(0, dtype=bool)
+        thr = self._conf_thres_array[cls_ids]
+        keep = scores >= thr
+        for c in np.unique(cls_ids):
+            b = float(self._bonus_array[c])
+            if b <= 0.0:
                 continue
-
-            if confidence < CLASS_CONF_LIST[class_id]:
+            cm = cls_ids == c
+            if keep[cm].any():
                 continue
-
-            x1 = (x1 - pad_left) / scale
-            x2 = (x2 - pad_left) / scale
-            y1 = (y1 - pad_top) / scale
-            y2 = (y2 - pad_top) / scale
-
-            x1 = np.clip(x1, 0, original_w - 1)
-            x2 = np.clip(x2, 0, original_w - 1)
-            y1 = np.clip(y1, 0, original_h - 1)
-            y2 = np.clip(y2, 0, original_h - 1)
-
-            detections.append(
-                {
-                    "x1": float(x1),
-                    "y1": float(y1),
-                    "x2": float(x2),
-                    "y2": float(y2),
-                    "cls_id": class_id,
-                    "conf": confidence,
-                }
+            idx = np.where(cm)[0]
+            top = int(idx[int(np.argmax(scores[idx]))])
+            if scores[top] >= self._conf_thres_array[c] - b:
+                keep[top] = True
+        return keep
+
+    def _per_view_pipeline(self, boxes: np.ndarray, scores: np.ndarray,
+                           cls_ids: np.ndarray
+                           ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        if len(boxes) > 1:
+            keep = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
+            boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
+        if len(scores) > self.max_det:
+            top = np.argsort(-scores)[: self.max_det]
+            boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]
+        if len(boxes) > 1:
+            boxes, scores, cls_ids = self._cross_class_dedup_op(
+                boxes, scores, cls_ids, self.cross_iou_thresh
             )
-
-        return detections
-
-    def predict(
-        self,
-        image: np.ndarray,
-        confidence_threshold: float = 0.1,
-        iou_threshold: float = 0.7,
-    ) -> list[dict]:
-        original_h, original_w = image.shape[:2]
-
-        x, scale, pad_left, pad_top = self.preprocess(image)
-
-        outputs = self.session.run(
-            self.output_names,
-            {self.input_name: x},
-        )
-
-        return self.postprocess_raw_yolo(
-            output=outputs[0],
-            original_shape=(original_h, original_w),
-            scale=scale,
-            pad_left=pad_left,
-            pad_top=pad_top,
-            conf_threshold=confidence_threshold,
-            iou_threshold=iou_threshold,
-        )
-
-# =========================
-# BOX UTILS
-# =========================
-
-def calculate_iou(box1: dict, box2: dict) -> float:
-    x1 = max(box1["x1"], box2["x1"])
-    y1 = max(box1["y1"], box2["y1"])
-    x2 = min(box1["x2"], box2["x2"])
-    y2 = min(box1["y2"], box2["y2"])
-
-    iw = max(0.0, x2 - x1)
-    ih = max(0.0, y2 - y1)
-    inter = iw * ih
-
-    a1 = max(0.0, box1["x2"] - box1["x1"]) * max(0.0, box1["y2"] - box1["y1"])
-    a2 = max(0.0, box2["x2"] - box2["x1"]) * max(0.0, box2["y2"] - box2["y1"])
-
-    union = a1 + a2 - inter
-    if union <= 0:
-        return 0.0
-
-    return inter / union
-
-
-def nms_same_class(
-    detections: list[dict],
-    iou_threshold: float = 0.7,
-) -> list[dict]:
-    if not detections:
-        return []
-
-    final = []
-
-    class_ids = sorted({box["cls_id"] for box in detections})
-
-    for class_id in class_ids:
-        boxes = [box for box in detections if box["cls_id"] == class_id]
-        boxes.sort(key=lambda box: box["conf"], reverse=True)
-
-        while boxes:
-            best = boxes.pop(0)
-            final.append(best)
-
-            boxes = [
-                box
-                for box in boxes
-                if calculate_iou(best, box) < iou_threshold
-            ]
-
-    final.sort(key=lambda box: box["conf"], reverse=True)
-    return final
-
-
-def iou_xyxy_to_many(box: np.ndarray, boxes: np.ndarray) -> np.ndarray:
-    x1 = np.maximum(box[0], boxes[:, 0])
-    y1 = np.maximum(box[1], boxes[:, 1])
-    x2 = np.minimum(box[2], boxes[:, 2])
-    y2 = np.minimum(box[3], boxes[:, 3])
-
-    inter_w = np.maximum(0.0, x2 - x1)
-    inter_h = np.maximum(0.0, y2 - y1)
-    inter_area = inter_w * inter_h
-
-    box_area = max(0.0, box[2] - box[0]) * max(0.0, box[3] - box[1])
-
-    boxes_area = np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) * np.maximum(
-        0.0,
-        boxes[:, 3] - boxes[:, 1],
-    )
-
-    union = box_area + boxes_area - inter_area
-    return inter_area / np.maximum(union, 1e-6)
-
-
-def fuse_cluster_numpy(cluster: np.ndarray) -> dict:
-    weights = np.maximum(cluster[:, 4], 1e-6)
-
-    fused_xyxy = np.average(
-        cluster[:, :4],
-        axis=0,
-        weights=weights,
-    )
-
-    best_idx = int(np.argmax(cluster[:, 4]))
-    best_box = cluster[best_idx]
-
-    return {
-        "x1": int(round(float(fused_xyxy[0]))),
-        "y1": int(round(float(fused_xyxy[1]))),
-        "x2": int(round(float(fused_xyxy[2]))),
-        "y2": int(round(float(fused_xyxy[3]))),
-        "cls_id": int(best_box[5]),
-        "conf": float(np.max(cluster[:, 4])),
-        "merged_count": int(len(cluster)),
-    }
-
-
-def merge_same_class_boxes(
-    detections: list[dict],
-    iou_threshold: float = 0.45,
-) -> list[dict]:
-    if not detections:
-        return []
-
-    boxes_array = np.array(
-        [
-            [
-                box["x1"],
-                box["y1"],
-                box["x2"],
-                box["y2"],
-                box["conf"],
-                box["cls_id"],
-            ]
-            for box in detections
-        ],
-        dtype=np.float32,
-    )
-
-    merged_detections = []
-
-    class_ids = np.unique(boxes_array[:, 5].astype(np.int32))
-
-    for class_id in class_ids:
-        class_boxes = boxes_array[boxes_array[:, 5] == class_id]
-
-        order = np.argsort(-class_boxes[:, 4])
-        class_boxes = class_boxes[order]
-
-        alive = np.ones(len(class_boxes), dtype=bool)
-
-        while np.any(alive):
-            seed_idx = int(np.flatnonzero(alive)[0])
-
-            cluster_indices = [seed_idx]
-            alive[seed_idx] = False
-
-            changed = True
-
-            while changed and np.any(alive):
-                changed = False
-
-                cluster = class_boxes[cluster_indices]
-                weights = np.maximum(cluster[:, 4], 1e-6)
-
-                current_fused_xyxy = np.average(
-                    cluster[:, :4],
-                    axis=0,
-                    weights=weights,
-                )
-
-                candidate_indices = np.flatnonzero(alive)
-                candidate_boxes = class_boxes[candidate_indices, :4]
-
-                ious = iou_xyxy_to_many(
-                    current_fused_xyxy,
-                    candidate_boxes,
+        return boxes, scores, cls_ids
+
+    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)
+
+        # Remap model cls_ids -> output cls_ids BEFORE the conf filter, so the
+        # per-class threshold/bonus arrays (indexed in `class_names` order) are
+        # applied to the right class.
+        n_model_cls = len(self.model_class_names)
+        vmask = (cls_ids >= 0) & (cls_ids < n_model_cls)
+        boxes, scores, cls_ids = boxes[vmask], scores[vmask], cls_ids[vmask]
+        cls_ids = self.cls_remap[cls_ids]
+
+        keep = self._conf_filter_mask(scores, cls_ids)
+        boxes = boxes[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+        if len(boxes) == 0:
+            return []
+
+        pad_w, pad_h = pad
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+
+        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
+        return self._build_results(boxes, scores, cls_ids)
+
+    def _decode_raw_yolo(self, preds: np.ndarray, ratio: float,
+                         pad: tuple[float, float],
+                         orig_size: tuple[int, int]) -> list[BoundingBox]:
+        if preds.ndim != 3 or preds.shape[0] != 1:
+            raise ValueError(f"Unexpected raw ONNX output shape: {preds.shape}")
+        preds = preds[0]
+        if preds.shape[0] <= 16 and preds.shape[1] > preds.shape[0]:
+            preds = preds.T
+        if preds.ndim != 2 or preds.shape[1] < 5:
+            raise ValueError(f"Unexpected raw output shape: {preds.shape}")
+
+        boxes_xywh = preds[:, :4].astype(np.float32)
+        cls_part = preds[:, 4:].astype(np.float32)
+        if cls_part.shape[1] == 1:
+            scores = cls_part[:, 0]
+            cls_ids = np.zeros(len(scores), dtype=np.int32)
+        else:
+            cls_ids = np.argmax(cls_part, axis=1).astype(np.int32)
+            scores = cls_part[np.arange(len(cls_part)), cls_ids]
+
+        # Remap model cls_ids -> output cls_ids BEFORE the conf filter, so the
+        # per-class threshold/bonus arrays (indexed in `class_names` order) are
+        # applied to the right class.
+        n_model_cls = len(self.model_class_names)
+        vmask = (cls_ids >= 0) & (cls_ids < n_model_cls)
+        boxes_xywh, scores, cls_ids = boxes_xywh[vmask], scores[vmask], cls_ids[vmask]
+        cls_ids = self.cls_remap[cls_ids]
+
+        keep = self._conf_filter_mask(scores, cls_ids)
+        boxes_xywh = boxes_xywh[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+        if len(boxes_xywh) == 0:
+            return []
+        boxes = self._xywh_to_xyxy(boxes_xywh)
+
+        pad_w, pad_h = pad
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+
+        boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
+        return self._build_results(boxes, scores, cls_ids)
+
+    @staticmethod
+    def _build_results(boxes: np.ndarray, scores: np.ndarray,
+                       cls_ids: np.ndarray) -> list[BoundingBox]:
+        results: list[BoundingBox] = []
+        for box, conf, cls_id in zip(boxes, scores, cls_ids):
+            x1, y1, x2, y2 = box.tolist()
+            if x2 <= x1 or y2 <= y1:
+                continue
+            results.append(
+                BoundingBox(
+                    x1=int(math.floor(x1)),
+                    y1=int(math.floor(y1)),
+                    x2=int(math.ceil(x2)),
+                    y2=int(math.ceil(y2)),
+                    cls_id=int(cls_id),
+                    conf=float(conf),
                 )
+            )
+        return results
 
-                matched_indices = candidate_indices[ious >= iou_threshold]
-
-                if len(matched_indices) > 0:
-                    cluster_indices.extend(matched_indices.tolist())
-                    alive[matched_indices] = False
-                    changed = True
-
-            final_cluster = class_boxes[cluster_indices]
-            merged_detections.append(fuse_cluster_numpy(final_cluster))
-
-    merged_detections.sort(
-        key=lambda box: box["conf"],
-        reverse=True,
-    )
-
-    return merged_detections
-
-
-# =========================
-# INFERENCE
-# =========================
-
-def reference_model(
-    image: np.ndarray,
-    model: ONNXYOLO,
-    confidence_threshold: float = 0.1,
-) -> list[dict]:
-    detections = model.predict(
-        image=image,
-        confidence_threshold=confidence_threshold,
-        iou_threshold=0.7,
-    )
-
-    return detections
-
-
-class BoundingBox(BaseModel):
-    x1: int
-    y1: int
-    x2: int
-    y2: int
-    cls_id: int
-    conf: float
-
-
-class Polygon(BaseModel):
-    cls_id: int
-    conf: float
-    points: list[tuple[int, int]]
-
-
-class TVFrameResult(BaseModel):
-    frame_id: int
-    boxes: list[BoundingBox] | None = None
-    polygons: list[Polygon] | None = None
-    keypoints: list[tuple[int, int]] | None = None
-
-
-class Miner:
-    """
-    This class is responsible for:
-    - Loading ML models.
-    - Running batched predictions on images.
-    - Parsing ML model outputs into structured results (TVFrameResult).
-
-    This class can be modified, but it must have the following to be compatible with the chute:
-        - be named `Miner`
-        - have a `predict_batch` function with the inputs and outputs specified
-        - be stored in a file called `miner.py` which lives in the root of the HFHub repo
-    """
-
-    def __init__(self, path_hf_repo: Path) -> None:
-        """
-        Loads all ML models from the repository.
-        -----(Adjust as needed)----
-
-        Args:
-            path_hf_repo (Path):
-                Path to the downloaded HuggingFace Hub repository
-
-        Returns:
-            None
-        """
-        self.model = ONNXYOLO(path_hf_repo / "weights.onnx", providers=["CPUExecutionProvider"])
-        print(f"✅ Model Loaded")
-
-        self._class_id_map = np.fromiter(
-            map(SUBNET_CLASS_NAMES.index, MODEL_CLASS_NAMES), dtype=np.int32,
-            count=len(MODEL_CLASS_NAMES),
-        )
+    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[2] != 3:
+            raise ValueError(f"Expected 3 channels, got shape={image.shape}")
+        if image.dtype != np.uint8:
+            image = image.astype(np.uint8)
+
+        input_tensor, ratio, pad, orig_size = self._preprocess(image)
+        expected = (1, 3, self.input_height, self.input_width)
+        if input_tensor.shape != expected:
+            raise ValueError(
+                f"Bad input tensor shape={input_tensor.shape}, expected={expected}"
+            )
 
-    def __repr__(self) -> str:
-        """
-        Information about miner returned in the health endpoint
-        to inspect the loaded ML models (and their types)
-        -----(Adjust as needed)----
-        """
-        return f"Model: {type(self.model).__name__}"
-
-    def predict_crime(self, image: np.ndarray, model: ONNXYOLO, confidence_threshold: float = 0.1) -> list[dict]:
-        detections = reference_model(
-            image=image,
-            model=model,
-            confidence_threshold=confidence_threshold,
+        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
+        return self._postprocess(outputs[0], ratio, pad, orig_size)
+
+    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        boxes_orig = self._predict_single(image)
+        flipped = cv2.flip(image, 1)
+        boxes_flip = self._predict_single(flipped)
+        w = image.shape[1]
+        boxes_flip = [
+            BoundingBox(
+                x1=w - b.x2, y1=b.y1, x2=w - b.x1, y2=b.y2,
+                cls_id=b.cls_id, conf=b.conf,
+            )
+            for b in boxes_flip
+        ]
+        all_boxes = boxes_orig + boxes_flip
+        if not all_boxes:
+            return []
+
+        coords = np.array(
+            [[b.x1, b.y1, b.x2, b.y2] for b in all_boxes], dtype=np.float32
         )
-
-        merged_boxes = merge_same_class_boxes(
-            detections,
-            iou_threshold=0.65,
+        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 []
+        if len(hard_keep) > self.max_det:
+            top = np.argsort(-scores[hard_keep])[: self.max_det]
+            hard_keep = hard_keep[top]
+        boosted = self._max_score_per_cluster(
+            coords[hard_keep], cls_ids[hard_keep],
+            coords, scores, cls_ids, self.iou_thres,
         )
 
-        return merged_boxes
-
-    def predict_batch(
-        self,
-        batch_images: list[ndarray],
-        offset: int,
-        n_keypoints: int,
-    ) -> list[TVFrameResult]:
-        """
-        Miner prediction for a batch of images.
-        Handles the orchestration of ML models and any preprocessing and postprocessing
-        -----(Adjust as needed)----
-
-        Args:
-            batch_images (list[np.ndarray]):
-                A list of images (as NumPy arrays) to process in this batch.
-            offset (int):
-                The frame number corresponding to the first image in the batch.
-                Used to correctly index frames in the output results.
-            n_keypoints (int):
-                The number of keypoints expected for each frame in this challenge type.
-
-        Returns:
-            list[TVFrameResult]:
-                A list of predictions for each image in the batch
-        """
-
-        bboxes: dict[int, list[BoundingBox]] = {}
-
-        for frame_number_in_batch, image in enumerate(batch_images):
-            merged_boxes = self.predict_crime(
-                image=image,
-                model=self.model,
-                confidence_threshold=0.1,
+        kept_coords = coords[hard_keep]
+        kept_cls = cls_ids[hard_keep]
+        if len(kept_coords) > 1:
+            kept_coords, boosted, kept_cls = self._cross_class_dedup_op(
+                kept_coords, boosted, kept_cls, self.cross_iou_thresh
             )
 
-            boxes = []
-            for box in merged_boxes:
-                x1, y1, x2, y2, model_cls_id, model_conf, _ = box.values()
-                bossted_conf = min(model_conf + BONUS_LIST[model_cls_id], 0.99)
-
-                sn_cls_id = self._class_id_map[int(model_cls_id)]
-
-                boxes.append(
-                    BoundingBox(
-                        x1=int(x1),
-                        y1=int(y1),
-                        x2=int(x2),
-                        y2=int(y2),
-                        cls_id=int(sn_cls_id),
-                        conf=float(bossted_conf),
-                    )
-                )
-            bboxes[offset + frame_number_in_batch] = boxes
-        print("✅ BBoxes predicted")
+        return [
+            BoundingBox(
+                x1=int(math.floor(kept_coords[j, 0])),
+                y1=int(math.floor(kept_coords[j, 1])),
+                x2=int(math.ceil(kept_coords[j, 2])),
+                y2=int(math.ceil(kept_coords[j, 3])),
+                cls_id=int(kept_cls[j]),
+                conf=float(boosted[j]),
+            )
+            for j in range(len(kept_coords))
+        ]
 
+    def predict_batch(self, batch_images: list[ndarray], offset: int,
+                      n_keypoints: int) -> list[TVFrameResult]:
         results: list[TVFrameResult] = []
-        for frame_number in range(offset, offset + len(batch_images)):
+        predict = self._predict_tta if self.use_tta else self._predict_single
+        for frame_number_in_batch, image in enumerate(batch_images):
+            try:
+                boxes = predict(image)
+            except Exception as e:
+                print(f"Inference failed for frame {offset + frame_number_in_batch}: {e}")
+                boxes = []
             results.append(
                 TVFrameResult(
-                    frame_id=frame_number,
-                    boxes=bboxes.get(frame_number, []),
+                    frame_id=offset + frame_number_in_batch,
+                    boxes=boxes,
+                    keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
                 )
             )
-        print("✅ Combined results as TVFrameResult")
-        return results
+        return results