Upload folder using huggingface_hub

miner.py

@@ -26,7 +26,7 @@ class TVFrameResult(BaseModel):
 class Miner:
     def __init__(self, path_hf_repo: Path) -> None:
         model_path = path_hf_repo / "weights.onnx"
-        self.class_names = ['person']
+        self.class_names = ["person"]
         print("ORT version:", ort.__version__)
 
         try:
@@ -67,21 +67,31 @@ class Miner:
         self.output_names = [output.name for output in self.session.get_outputs()]
         self.input_shape = self.session.get_inputs()[0].shape
 
-        # Your export is fixed-size 1280, but we still read actual ONNX input shape first.
         self.input_height = self._safe_dim(self.input_shape[2], default=1280)
         self.input_width = self._safe_dim(self.input_shape[3], default=1280)
 
-        # Tuned for validator scoring: reduce FP (FALSE_POSITIVE pillar),
-        # preserve recall (MAP50, RECALL), improve precision.
-        self.conf_thres = 0.2  # Higher = fewer FP, slightly lower recall
-        self.iou_thres = 0.5   # Lower = suppress duplicate detections (FP)
-        self.max_det = 100     # Cap detections; sports ~20-30 persons
+        # ---------- Scoring-oriented thresholds ----------
+        # Low threshold for candidate generation
+        self.conf_thres = 0.1
+
+        # High-confidence boxes can survive without TTA confirmation
+        self.conf_high = 0.5
+
+        # NMS threshold
+        self.iou_thres = 0.50
+
+        # TTA confirmation IoU
+        self.tta_match_iou = 0.55
+
+        self.max_det = 300
         self.use_tta = True
 
-        # Box sanity: filter tiny/spurious detections (common FP source)
-        self.min_box_area = 12 * 12  # ~144 px²
-        self.min_side = 8
-        self.max_aspect_ratio = 8.0
+        # Box sanity filters
+        self.min_box_area = 16 * 16
+        self.min_w = 6
+        self.min_h = 6
+        self.max_aspect_ratio = 6.0
+        self.max_box_area_ratio = 0.95
 
         print(f"✅ ONNX model loaded from: {model_path}")
         print(f"✅ ONNX providers: {self.session.get_providers()}")
@@ -103,13 +113,6 @@ 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
 
@@ -145,14 +148,6 @@ 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(
@@ -183,93 +178,56 @@ class Miner:
         out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
         return out
 
-    def _soft_nms(
-        self,
+    @staticmethod
+    def _hard_nms(
         boxes: np.ndarray,
         scores: np.ndarray,
-        sigma: float = 0.5,
-        score_thresh: float = 0.01,
-    ) -> tuple[np.ndarray, np.ndarray]:
-        """
-        Soft-NMS: Gaussian decay of overlapping scores instead of hard removal.
-        Returns (kept_original_indices, updated_scores).
-        """
-        N = len(boxes)
-        if N == 0:
-            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
-
-        boxes = boxes.astype(np.float32, copy=True)
-        scores = scores.astype(np.float32, copy=True)
-        order = np.arange(N)
+        iou_thresh: float,
+    ) -> np.ndarray:
+        if len(boxes) == 0:
+            return np.array([], dtype=np.intp)
 
-        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]]
+        boxes = np.asarray(boxes, dtype=np.float32)
+        scores = np.asarray(scores, dtype=np.float32)
+        order = np.argsort(scores)[::-1]
+        keep = []
 
-            if i + 1 >= N:
+        while len(order) > 0:
+            i = order[0]
+            keep.append(i)
+            if len(order) == 1:
                 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])
+            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, 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)
+            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]
 
-        mask = scores > score_thresh
-        return order[mask], scores[mask]
+        return np.array(keep, dtype=np.intp)
 
     @staticmethod
-    def _hard_nms(
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        """
-        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)
-        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
-            keep.append(idx)
-            bi = boxes[idx]
-            for k in range(i + 1, N):
-                jdx = order[k]
-                if suppressed[jdx]:
-                    continue
-                bj = boxes[jdx]
-                xx1 = max(bi[0], bj[0])
-                yy1 = max(bi[1], bj[1])
-                xx2 = min(bi[2], bj[2])
-                yy2 = min(bi[3], bj[3])
-                inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1)
-                area_i = (bi[2] - bi[0]) * (bi[3] - bi[1])
-                area_j = (bj[2] - bj[0]) * (bj[3] - bj[1])
-                iou = inter / (area_i + area_j - inter + 1e-7)
-                if iou > iou_thresh:
-                    suppressed[jdx] = True
-        return np.array(keep)
+    def _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,
@@ -278,69 +236,44 @@ class Miner:
         cls_ids: np.ndarray,
         orig_size: tuple[int, int],
     ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """Filter out tiny, degenerate, or implausible boxes (common FP)."""
         if len(boxes) == 0:
             return boxes, scores, cls_ids
+
         orig_w, orig_h = orig_size
         image_area = float(orig_w * orig_h)
+
         keep = []
         for i, box in enumerate(boxes):
             x1, y1, x2, y2 = box.tolist()
             bw = x2 - x1
             bh = y2 - y1
+
             if bw <= 0 or bh <= 0:
                 continue
-            if bw < self.min_side or bh < self.min_side:
+            if bw < self.min_w or bh < self.min_h:
                 continue
+
             area = bw * bh
             if area < self.min_box_area:
                 continue
-            if area > 0.95 * image_area:
+            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),
             )
-        k = np.array(keep, dtype=np.intp)
-        return boxes[k], scores[k], cls_ids[k]
 
-    @staticmethod
-    def _max_score_per_cluster(
-        coords: np.ndarray,
-        scores: np.ndarray,
-        keep_indices: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        """
-        For each kept box, return the max original score among itself and any
-        box that overlaps it with IOU >= iou_thresh (so TTA cluster keeps best conf).
-        """
-        n_keep = len(keep_indices)
-        if n_keep == 0:
-            return np.array([], dtype=np.float32)
-        out = np.empty(n_keep, dtype=np.float32)
-        coords = np.asarray(coords, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
-        for i in range(n_keep):
-            idx = keep_indices[i]
-            bi = coords[idx]
-            xx1 = np.maximum(bi[0], coords[:, 0])
-            yy1 = np.maximum(bi[1], coords[:, 1])
-            xx2 = np.minimum(bi[2], coords[:, 2])
-            yy2 = np.minimum(bi[3], coords[:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            area_i = (bi[2] - bi[0]) * (bi[3] - bi[1])
-            areas_j = (coords[:, 2] - coords[:, 0]) * (coords[:, 3] - coords[:, 1])
-            iou = inter / (area_i + areas_j - inter + 1e-7)
-            in_cluster = iou >= iou_thresh
-            out[i] = float(np.max(scores[in_cluster]))
-        return out
+        keep = np.array(keep, dtype=np.intp)
+        return boxes[keep], scores[keep], cls_ids[keep]
 
     def _decode_final_dets(
         self,
@@ -348,13 +281,7 @@ 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]
 
@@ -365,6 +292,13 @@ class Miner:
         scores = preds[:, 4].astype(np.float32)
         cls_ids = preds[:, 5].astype(np.int32)
 
+        # person only
+        keep = cls_ids == 0
+        boxes = boxes[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+
+        # candidate threshold
         keep = scores >= self.conf_thres
         boxes = boxes[keep]
         scores = scores[keep]
@@ -376,51 +310,34 @@ 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))
 
-        # Box sanity filter (reduces FP)
-        boxes, scores, cls_ids = self._filter_sane_boxes(
-            boxes, scores, cls_ids, orig_size
-        )
+        boxes, scores, cls_ids = self._filter_sane_boxes(boxes, scores, cls_ids, orig_size)
         if len(boxes) == 0:
             return []
 
-        # NMS to remove duplicates (model may output overlapping boxes)
-        if len(boxes) > 1:
-            if apply_optional_dedup:
-                keep_idx, scores = self._soft_nms(boxes, scores)
-                boxes = boxes[keep_idx]
-                cls_ids = cls_ids[keep_idx]
-            else:
-                keep_idx = self._hard_nms(boxes, scores, self.iou_thres)
-                keep_idx = keep_idx[: self.max_det]
-                boxes = boxes[keep_idx]
-                scores = scores[keep_idx]
-                cls_ids = cls_ids[keep_idx]
-
-        results: list[BoundingBox] = []
-        for box, conf, cls_id in zip(boxes, scores, cls_ids):
-            x1, y1, x2, y2 = box.tolist()
+        keep_idx = self._hard_nms(boxes, scores, self.iou_thres)
+        keep_idx = keep_idx[: self.max_det]
 
-            if x2 <= x1 or y2 <= y1:
-                continue
+        boxes = boxes[keep_idx]
+        scores = scores[keep_idx]
+        cls_ids = cls_ids[keep_idx]
 
-            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 [
+            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),
             )
-
-        return results
+            for box, conf, cls_id in zip(boxes, scores, cls_ids)
+            if box[2] > box[0] and box[3] > box[1]
+        ]
 
     def _decode_raw_yolo(
         self,
@@ -429,15 +346,8 @@ 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}")
 
@@ -451,14 +361,31 @@ class Miner:
             raise ValueError(f"Unexpected normalized raw output shape: {preds.shape}")
 
         boxes_xywh = preds[:, :4].astype(np.float32)
-        cls_part = preds[:, 4:].astype(np.float32)
-
-        if cls_part.shape[1] == 1:
-            scores = cls_part[:, 0]
+        tail = preds[:, 4:].astype(np.float32)
+
+        # Supports:
+        # [x,y,w,h,score]                 single-class
+        # [x,y,w,h,obj,cls]               YOLO standard single-class
+        # [x,y,w,h,obj,cls1,cls2,...]     multi-class
+        if tail.shape[1] == 1:
+            scores = tail[:, 0]
+            cls_ids = np.zeros(len(scores), dtype=np.int32)
+        elif tail.shape[1] == 2:
+            obj = tail[:, 0]
+            cls_prob = tail[:, 1]
+            scores = obj * cls_prob
             cls_ids = np.zeros(len(scores), dtype=np.int32)
         else:
-            cls_ids = np.argmax(cls_part, axis=1).astype(np.int32)
-            scores = cls_part[np.arange(len(cls_part)), cls_ids]
+            obj = tail[:, 0]
+            class_probs = tail[:, 1:]
+            cls_ids = np.argmax(class_probs, axis=1).astype(np.int32)
+            cls_scores = class_probs[np.arange(len(class_probs)), cls_ids]
+            scores = obj * cls_scores
+
+        keep = cls_ids == 0
+        boxes_xywh = boxes_xywh[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
 
         keep = scores >= self.conf_thres
         boxes_xywh = boxes_xywh[keep]
@@ -470,12 +397,6 @@ class Miner:
 
         boxes = self._xywh_to_xyxy(boxes_xywh)
 
-        keep_idx = self._hard_nms(boxes, scores, self.iou_thres)
-        keep_idx = keep_idx[: self.max_det]
-        boxes = boxes[keep_idx]
-        scores = scores[keep_idx]
-        cls_ids = cls_ids[keep_idx]
-
         pad_w, pad_h = pad
         orig_w, orig_h = orig_size
 
@@ -484,31 +405,29 @@ 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_w, orig_h)
-        )
+        boxes, scores, cls_ids = self._filter_sane_boxes(boxes, scores, cls_ids, orig_size)
         if len(boxes) == 0:
             return []
 
-        results: list[BoundingBox] = []
-        for box, conf, cls_id in zip(boxes, scores, cls_ids):
-            x1, y1, x2, y2 = box.tolist()
+        keep_idx = self._hard_nms(boxes, scores, self.iou_thres)
+        keep_idx = keep_idx[: self.max_det]
 
-            if x2 <= x1 or y2 <= y1:
-                continue
+        boxes = boxes[keep_idx]
+        scores = scores[keep_idx]
+        cls_ids = cls_ids[keep_idx]
 
-            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 [
+            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),
             )
-
-        return results
+            for box, conf, cls_id in zip(boxes, scores, cls_ids)
+            if box[2] > box[0] and box[3] > box[1]
+        ]
 
     def _postprocess(
         self,
@@ -517,19 +436,12 @@ 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)
 
-        # final detections: [1,N,6]
-        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 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]:
@@ -559,58 +471,104 @@ class Miner:
         det_output = outputs[0]
         return self._postprocess(det_output, ratio, pad, orig_size)
 
-    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+    def _merge_tta_consensus(
+        self,
+        boxes_orig: list[BoundingBox],
+        boxes_flip: list[BoundingBox],
+    ) -> list[BoundingBox]:
         """
-        Horizontal-flip TTA: merge original + flipped via hard NMS.
-        Boost confidence for consensus detections (both views agree) to improve
-        mAP: validator sorts by confidence, so higher conf for TP helps PR curve.
+        Keep:
+        - any box with conf >= conf_high
+        - low/medium-conf boxes only if confirmed across TTA views
+        Then run final hard NMS.
         """
-        boxes_orig = self._predict_single(image)
+        if not boxes_orig and not boxes_flip:
+            return []
 
-        flipped = cv2.flip(image, 1)
-        boxes_flip = self._predict_single(flipped)
+        coords_o = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty((0, 4), dtype=np.float32)
+        scores_o = np.array([b.conf for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty((0,), dtype=np.float32)
+
+        coords_f = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty((0, 4), dtype=np.float32)
+        scores_f = np.array([b.conf for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty((0,), dtype=np.float32)
+
+        accepted_boxes = []
+        accepted_scores = []
+
+        # Original view candidates
+        for i in range(len(coords_o)):
+            score = scores_o[i]
+            if score >= self.conf_high:
+                accepted_boxes.append(coords_o[i])
+                accepted_scores.append(score)
+            elif len(coords_f) > 0:
+                ious = self._box_iou_one_to_many(coords_o[i], coords_f)
+                j = int(np.argmax(ious))
+                if ious[j] >= self.tta_match_iou:
+                    fused_score = max(score, scores_f[j])
+                    accepted_boxes.append(coords_o[i])
+                    accepted_scores.append(fused_score)
+
+        # Flipped-view high-confidence boxes that original missed
+        for i in range(len(coords_f)):
+            score = scores_f[i]
+            if score < self.conf_high:
+                continue
 
-        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
-        ]
+            if len(coords_o) == 0:
+                accepted_boxes.append(coords_f[i])
+                accepted_scores.append(score)
+                continue
+
+            ious = self._box_iou_one_to_many(coords_f[i], coords_o)
+            if np.max(ious) < self.tta_match_iou:
+                accepted_boxes.append(coords_f[i])
+                accepted_scores.append(score)
 
-        all_boxes = boxes_orig + boxes_flip
-        if len(all_boxes) == 0:
+        if not accepted_boxes:
             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)
+        boxes = np.array(accepted_boxes, dtype=np.float32)
+        scores = np.array(accepted_scores, dtype=np.float32)
 
-        hard_keep = self._hard_nms(coords, scores, self.iou_thres)
-        if len(hard_keep) == 0:
-            return []
+        keep = self._hard_nms(boxes, scores, self.iou_thres)
+        keep = keep[: self.max_det]
 
-        hard_keep = hard_keep[: self.max_det]
+        out = []
+        for idx in keep:
+            x1, y1, x2, y2 = boxes[idx].tolist()
+            out.append(
+                BoundingBox(
+                    x1=int(math.floor(x1)),
+                    y1=int(math.floor(y1)),
+                    x2=int(math.ceil(x2)),
+                    y2=int(math.ceil(y2)),
+                    cls_id=0,
+                    conf=float(scores[idx]),
+                )
+            )
+        return out
 
-        # Boost confidence when both views agree (overlapping detections)
-        boosted = self._max_score_per_cluster(
-            coords, scores, hard_keep, self.iou_thres
-        )
+    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        boxes_orig = self._predict_single(image)
 
-        return [
+        flipped = cv2.flip(image, 1)
+        boxes_flip_raw = self._predict_single(flipped)
+
+        w = image.shape[1]
+        boxes_flip = [
             BoundingBox(
-                x1=all_boxes[i].x1,
-                y1=all_boxes[i].y1,
-                x2=all_boxes[i].x2,
-                y2=all_boxes[i].y2,
-                cls_id=all_boxes[i].cls_id,
-                conf=float(boosted[j]),
+                x1=w - b.x2,
+                y1=b.y1,
+                x2=w - b.x1,
+                y2=b.y2,
+                cls_id=b.cls_id,
+                conf=b.conf,
             )
-            for j, i in enumerate(hard_keep)
+            for b in boxes_flip_raw
         ]
 
+        return self._merge_tta_consensus(boxes_orig, boxes_flip)
+
     def predict_batch(
         self,
         batch_images: list[ndarray],

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:be6e4b7e468f655482d0f73509954721601eacb68d376d8191a14bdb7f3d3105
+oid sha256:4a7fac9f674924560d7a4788e2e1a4733b1055e4380f9e72e0720ed9fb3d155b
 size 19404973