scorevision: push artifact

miner.py

@@ -22,13 +22,22 @@ class TVFrameResult(BaseModel):
     boxes: list[BoundingBox]
     keypoints: list[tuple[int, int]]
 
+SIZE = 1280
+
 
 class Miner:
-    def __init__(self,
-        path_hf_repo: Path
-    ) -> None:
+    def __init__(self, path_hf_repo: Path) -> None:
         model_path = path_hf_repo / "weights.onnx"
-        self.class_names = ['bus', 'car', 'truck', 'motorcycle']
+        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 = ["person"]
         print("ORT version:", ort.__version__)
 
         try:
@@ -69,32 +78,14 @@ class Miner:
         self.output_names = [output.name for output in self.session.get_outputs()]
         self.input_shape = self.session.get_inputs()[0].shape
 
-        self.input_height = self._safe_dim(self.input_shape[2], default=1280)
-        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
-
-        # ---------- Scoring-oriented thresholds ----------
-        # Low threshold for candidate generation
-        self.conf_thres = 0.15
-
-        # High-confidence boxes can survive without TTA confirmation
-        self.conf_high = 0.50
-
-        # NMS threshold
-        self.iou_thres = 0.66
+        self.input_height = self._safe_dim(self.input_shape[2], default=SIZE)
+        self.input_width = self._safe_dim(self.input_shape[3], default=SIZE)
 
-        # TTA confirmation IoU
-        self.tta_match_iou = 0.91
-
-        self.max_det = 150
+        self.conf_thres = 0.45
+        self.iou_thres = 0.5
+        self.max_det = 30
         self.use_tta = True
 
-        # Box sanity filters
-        self.min_box_area = 4 * 4
-        self.min_w = 2
-        self.min_h = 2
-        self.max_aspect_ratio = 12.0
-        self.max_box_area_ratio = 0.95
-
         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}")
@@ -115,6 +106,13 @@ class Miner:
         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
 
@@ -150,6 +148,14 @@ class Miner:
     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(
@@ -180,125 +186,125 @@ class Miner:
         out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
         return out
 
-    @staticmethod
-    def _hard_nms(
+    def _soft_nms(
+        self,
         boxes: np.ndarray,
         scores: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        if len(boxes) == 0:
-            return np.array([], dtype=np.intp)
-
-        boxes = np.asarray(boxes, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
-        order = np.argsort(scores)[::-1]
-        keep = []
+        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.
+        Returns (kept_original_indices, updated_scores).
+        """
+        N = len(boxes)
+        if N == 0:
+            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
 
-        while len(order) > 0:
-            i = order[0]
-            keep.append(i)
-            if len(order) == 1:
-                break
+        boxes = boxes.astype(np.float32, copy=True)
+        scores = scores.astype(np.float32, copy=True)
+        order = np.arange(N)
 
-            rest = order[1:]
+        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]]
 
-            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])
+            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 = np.maximum(0.0, (boxes[i, 2] - boxes[i, 0])) * np.maximum(0.0, (boxes[i, 3] - boxes[i, 1]))
-            area_r = np.maximum(0.0, (boxes[rest, 2] - boxes[rest, 0])) * np.maximum(0.0, (boxes[rest, 3] - boxes[rest, 1]))
-
-            iou = inter / (area_i + area_r - inter + 1e-7)
-            order = rest[iou <= iou_thresh]
+            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)
 
-        return np.array(keep, dtype=np.intp)
+        mask = scores > score_thresh
+        return order[mask], scores[mask]
 
-    @classmethod
-    def _nms_per_class(
-        cls,
+    @staticmethod
+    def _hard_nms(
         boxes: np.ndarray,
         scores: np.ndarray,
-        cls_ids: np.ndarray,
         iou_thresh: float,
-        max_det: int,
     ) -> np.ndarray:
-        """NMS within each class so overlapping car vs bus predictions are not merged away."""
-        if len(boxes) == 0:
+        """
+        Standard NMS: keep one box per overlapping cluster (the one with highest score).
+        Returns indices of kept boxes (into the boxes/scores arrays).
+        """
+        N = len(boxes)
+        if N == 0:
             return np.array([], dtype=np.intp)
-        keep_all: list[int] = []
-        for c in np.unique(cls_ids):
-            idxs = np.nonzero(cls_ids == c)[0]
-            if len(idxs) == 0:
+        boxes = np.asarray(boxes, dtype=np.float32)
+        scores = np.asarray(scores, dtype=np.float32)
+        order = np.argsort(scores)[::-1]
+        keep: list[int] = []
+        suppressed = np.zeros(N, dtype=bool)
+        for i in range(N):
+            idx = order[i]
+            if suppressed[idx]:
                 continue
-            local_keep = cls._hard_nms(boxes[idxs], scores[idxs], iou_thresh)
-            keep_all.extend(idxs[local_keep].tolist())
-        keep_all = np.array(keep_all, dtype=np.intp)
-        order = np.argsort(scores[keep_all])[::-1]
-        return keep_all[order[:max_det]]
+            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)
 
     @staticmethod
-    def _box_iou_one_to_many(box: np.ndarray, boxes: np.ndarray) -> np.ndarray:
-        xx1 = np.maximum(box[0], boxes[:, 0])
-        yy1 = np.maximum(box[1], boxes[:, 1])
-        xx2 = np.minimum(box[2], boxes[:, 2])
-        yy2 = np.minimum(box[3], boxes[:, 3])
-
-        inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-
-        area_a = max(0.0, (box[2] - box[0]) * (box[3] - box[1]))
-        area_b = np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) * np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
-
-        return inter / (area_a + area_b - inter + 1e-7)
-
-    def _filter_sane_boxes(
-        self,
-        boxes: np.ndarray,
+    def _max_score_per_cluster(
+        coords: np.ndarray,
         scores: np.ndarray,
-        cls_ids: np.ndarray,
-        orig_size: tuple[int, int],
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        if len(boxes) == 0:
-            return boxes, scores, cls_ids
-
-        orig_w, orig_h = orig_size
-        image_area = float(orig_w * orig_h)
-
-        keep = []
-        for i, box in enumerate(boxes):
-            x1, y1, x2, y2 = box.tolist()
-            bw = x2 - x1
-            bh = y2 - y1
-
-            if bw <= 0 or bh <= 0:
-                continue
-            if bw < self.min_w or bh < self.min_h:
-                continue
-
-            area = bw * bh
-            if area < self.min_box_area:
-                continue
-            if area > self.max_box_area_ratio * image_area:
-                continue
-
-            ar = max(bw / max(bh, 1e-6), bh / max(bw, 1e-6))
-            if ar > self.max_aspect_ratio:
-                continue
-
-            keep.append(i)
-
-        if not keep:
-            return (
-                np.empty((0, 4), dtype=np.float32),
-                np.empty((0,), dtype=np.float32),
-                np.empty((0,), dtype=np.int32),
-            )
-
-        keep = np.array(keep, dtype=np.intp)
-        return boxes[keep], scores[keep], cls_ids[keep]
+        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,
@@ -306,7 +312,13 @@ class Miner:
         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]
 
@@ -317,9 +329,6 @@ class Miner:
         scores = preds[:, 4].astype(np.float32)
         cls_ids = preds[:, 5].astype(np.int32)
 
-        # All trained vehicle classes pass: bus, car, truck, motorcycle.
-
-        # candidate threshold
         keep = scores >= self.conf_thres
         boxes = boxes[keep]
         scores = scores[keep]
@@ -331,35 +340,36 @@ class Miner:
         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))
 
-        boxes, scores, cls_ids = self._filter_sane_boxes(boxes, scores, cls_ids, orig_size)
-        if len(boxes) == 0:
-            return []
+        if apply_optional_dedup and len(boxes) > 1:
+            keep_idx, scores = self._soft_nms(boxes, scores)
+            boxes = boxes[keep_idx]
+            cls_ids = cls_ids[keep_idx]
 
-        keep_idx = self._nms_per_class(
-            boxes, scores, cls_ids, self.iou_thres, self.max_det
-        )
+        results: list[BoundingBox] = []
+        for box, conf, cls_id in zip(boxes, scores, cls_ids):
+            x1, y1, x2, y2 = box.tolist()
 
-        boxes = boxes[keep_idx]
-        scores = scores[keep_idx]
-        cls_ids = cls_ids[keep_idx]
+            if x2 <= x1 or y2 <= y1:
+                continue
 
-        return [
-            BoundingBox(
-                x1=int(math.floor(box[0])),
-                y1=int(math.floor(box[1])),
-                x2=int(math.ceil(box[2])),
-                y2=int(math.ceil(box[3])),
-                cls_id=int(cls_id),
-                conf=float(conf),
+            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),
+                )
             )
-            for box, conf, cls_id in zip(boxes, scores, cls_ids)
-            if box[2] > box[0] and box[3] > box[1]
-        ]
+
+        return results
 
     def _decode_raw_yolo(
         self,
@@ -368,8 +378,15 @@ class Miner:
         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}")
 
@@ -383,26 +400,14 @@ class Miner:
             raise ValueError(f"Unexpected normalized raw output shape: {preds.shape}")
 
         boxes_xywh = preds[:, :4].astype(np.float32)
-        tail = preds[:, 4:].astype(np.float32)
-
-        # Supports:
-        # [x,y,w,h,score]                 single-class
-        # [x,y,w,h,obj,cls]               YOLO standard single-class
-        # [x,y,w,h,obj,cls1,cls2,...]     multi-class
-        if tail.shape[1] == 1:
-            scores = tail[:, 0]
-            cls_ids = np.zeros(len(scores), dtype=np.int32)
-        elif tail.shape[1] == 2:
-            obj = tail[:, 0]
-            cls_prob = tail[:, 1]
-            scores = obj * cls_prob
+        cls_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:
-            obj = tail[:, 0]
-            class_probs = tail[:, 1:]
-            cls_ids = np.argmax(class_probs, axis=1).astype(np.int32)
-            cls_scores = class_probs[np.arange(len(class_probs)), cls_ids]
-            scores = obj * cls_scores
+            cls_ids = np.argmax(cls_part, axis=1).astype(np.int32)
+            scores = cls_part[np.arange(len(cls_part)), cls_ids]
 
         keep = scores >= self.conf_thres
         boxes_xywh = boxes_xywh[keep]
@@ -413,6 +418,12 @@ class Miner:
             return []
 
         boxes = self._xywh_to_xyxy(boxes_xywh)
+        keep_idx, scores = self._soft_nms(boxes, scores)
+        keep_idx = keep_idx[: self.max_det]
+        scores = scores[: self.max_det]
+
+        boxes = boxes[keep_idx]
+        cls_ids = cls_ids[keep_idx]
 
         pad_w, pad_h = pad
         orig_w, orig_h = orig_size
@@ -422,30 +433,25 @@ class Miner:
         boxes /= ratio
         boxes = self._clip_boxes(boxes, (orig_w, orig_h))
 
-        boxes, scores, cls_ids = self._filter_sane_boxes(boxes, scores, cls_ids, orig_size)
-        if len(boxes) == 0:
-            return []
-
-        keep_idx = self._nms_per_class(
-            boxes, scores, cls_ids, self.iou_thres, self.max_det
-        )
+        results: list[BoundingBox] = []
+        for box, conf, cls_id in zip(boxes, scores, cls_ids):
+            x1, y1, x2, y2 = box.tolist()
 
-        boxes = boxes[keep_idx]
-        scores = scores[keep_idx]
-        cls_ids = cls_ids[keep_idx]
+            if x2 <= x1 or y2 <= y1:
+                continue
 
-        return [
-            BoundingBox(
-                x1=int(math.floor(box[0])),
-                y1=int(math.floor(box[1])),
-                x2=int(math.ceil(box[2])),
-                y2=int(math.ceil(box[3])),
-                cls_id=int(cls_id),
-                conf=float(conf),
+            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),
+                )
             )
-            for box, conf, cls_id in zip(boxes, scores, cls_ids)
-            if box[2] > box[0] and box[3] > box[1]
-        ]
+
+        return results
 
     def _postprocess(
         self,
@@ -454,12 +460,19 @@ class Miner:
         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)
 
-        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] >= 6:
+        # 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]:
@@ -489,110 +502,49 @@ class Miner:
         det_output = outputs[0]
         return self._postprocess(det_output, ratio, pad, orig_size)
 
-    def _merge_tta_consensus(
-        self,
-        boxes_orig: list[BoundingBox],
-        boxes_flip: list[BoundingBox],
-    ) -> list[BoundingBox]:
-        """
-        Keep:
-        - any box with conf >= conf_high
-        - low/medium-conf boxes only if confirmed across TTA views
-        Then run final hard NMS.
-        """
-        if not boxes_orig and not boxes_flip:
-            return []
-
-        coords_o = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty((0, 4), dtype=np.float32)
-        scores_o = np.array([b.conf for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty((0,), dtype=np.float32)
-        cls_o = np.array([b.cls_id for b in boxes_orig], dtype=np.int32) if boxes_orig else np.empty((0,), dtype=np.int32)
-
-        coords_f = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty((0, 4), dtype=np.float32)
-        scores_f = np.array([b.conf for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty((0,), dtype=np.float32)
-        cls_f = np.array([b.cls_id for b in boxes_flip], dtype=np.int32) if boxes_flip else np.empty((0,), dtype=np.int32)
-
-        accepted_boxes = []
-        accepted_scores = []
-        accepted_cls = []
-
-        # Original view candidates
-        for i in range(len(coords_o)):
-            score = scores_o[i]
-            if score >= self.conf_high:
-                accepted_boxes.append(coords_o[i])
-                accepted_scores.append(score)
-                accepted_cls.append(int(cls_o[i]))
-            elif len(coords_f) > 0:
-                ious = self._box_iou_one_to_many(coords_o[i], coords_f)
-                j = int(np.argmax(ious))
-                if ious[j] >= self.tta_match_iou:
-                    fused_score = max(score, scores_f[j])
-                    accepted_boxes.append(coords_o[i])
-                    accepted_scores.append(fused_score)
-                    accepted_cls.append(int(cls_o[i]))
-
-        # Flipped-view high-confidence boxes that original missed
-        for i in range(len(coords_f)):
-            score = scores_f[i]
-            if score < self.conf_high:
-                continue
-
-            if len(coords_o) == 0:
-                accepted_boxes.append(coords_f[i])
-                accepted_scores.append(score)
-                accepted_cls.append(int(cls_f[i]))
-                continue
-
-            ious = self._box_iou_one_to_many(coords_f[i], coords_o)
-            if np.max(ious) < self.tta_match_iou:
-                accepted_boxes.append(coords_f[i])
-                accepted_scores.append(score)
-                accepted_cls.append(int(cls_f[i]))
-
-        if not accepted_boxes:
-            return []
-
-        boxes = np.array(accepted_boxes, dtype=np.float32)
-        scores = np.array(accepted_scores, dtype=np.float32)
-        cls_ids = np.array(accepted_cls, dtype=np.int32)
-
-        keep = self._nms_per_class(boxes, scores, cls_ids, self.iou_thres, self.max_det)
-
-        out = []
-        for idx in keep:
-            x1, y1, x2, y2 = boxes[idx].tolist()
-            out.append(
-                BoundingBox(
-                    x1=int(math.floor(x1)),
-                    y1=int(math.floor(y1)),
-                    x2=int(math.ceil(x2)),
-                    y2=int(math.ceil(y2)),
-                    cls_id=int(cls_ids[idx]),
-                    conf=float(scores[idx]),
-                )
-            )
-        return out
-
     def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        """Horizontal-flip TTA: merge original + flipped via hard NMS."""
         boxes_orig = self._predict_single(image)
 
         flipped = cv2.flip(image, 1)
-        boxes_flip_raw = self._predict_single(flipped)
+        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,
+                x1=w - b.x2, y1=b.y1, x2=w - b.x1, y2=b.y2,
+                cls_id=b.cls_id, conf=b.conf,
             )
-            for b in boxes_flip_raw
+            for b in boxes_flip
         ]
 
-        return self._merge_tta_consensus(boxes_orig, 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)
+
+        hard_keep = self._hard_nms(coords, scores, self.iou_thres)
+        if len(hard_keep) == 0:
+            return []
+
+        # _hard_nms already orders kept indices by descending score.
+        hard_keep = hard_keep[: self.max_det]
+
+        return [
+            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(scores[i]),
+            )
+            for i in hard_keep
+        ]
 
     def predict_batch(
         self,
@@ -611,7 +563,14 @@ class Miner:
             except Exception as e:
                 print(f"⚠️ Inference failed for frame {offset + frame_number_in_batch}: {e}")
                 boxes = []
-
+            # for box in boxes:
+            #     if box.cls_id == 2:
+            #         box.cls_id = 3
+            #     elif box.cls_id == 3:
+            #         box.cls_id = 2
+                    
+            
+            
             results.append(
                 TVFrameResult(
                     frame_id=offset + frame_number_in_batch,
@@ -621,3 +580,53 @@ class Miner:
             )
 
         return results
+    
+
+if __name__ == "__main__":
+    # Simple manual test: load weights.onnx, run on 1.png, and draw bboxes
+    repo_dir = Path(__file__).parent
+    miner = Miner(repo_dir)
+
+    image_path = repo_dir / "car1.png"
+    if not image_path.exists():
+        raise FileNotFoundError(f"Test image not found: {image_path}")
+
+    image = cv2.imread(str(image_path), cv2.IMREAD_COLOR)
+    if image is None:
+        raise RuntimeError(f"Failed to read image: {image_path}")
+
+    results = miner.predict_batch([image], offset=0, n_keypoints=0)
+    # Draw bounding boxes on a copy of the image
+    vis = image.copy()
+    colors = [(0, 255, 0), (0, 0, 255), (255, 0, 0)]
+    for frame in results:
+        print(f"Frame {frame.frame_id}:")
+        for i, box in enumerate(frame.boxes):
+            color = colors[i % len(colors)]
+            cv2.rectangle(
+                vis,
+                (box.x1, box.y1),
+                (box.x2, box.y2),
+                color,
+                2,
+            )
+            label = f"{box.cls_id }_{miner.class_names[box.cls_id] if box.cls_id < len(miner.class_names) else box.cls_id}:{box.conf:.2f}"
+            cv2.putText(
+                vis,
+                label,
+                (box.x1, max(0, box.y1 - 5)),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                box.conf,
+                color,
+                1,
+                cv2.LINE_AA,
+            )
+            print(
+                f"  cls={box.cls_id} conf={box.conf:.3f} "
+                f"box=({box.x1},{box.y1},{box.x2},{box.y2})"
+            )
+        print(len(frame.boxes))
+
+    out_path = repo_dir / f"1_out_iou{miner.iou_thres:.2f}.png"
+    cv2.imwrite(str(out_path), vis)
+    print(f"Saved visualization to: {out_path}")