diff --git a/src/microwink/seg.py b/src/microwink/seg.py
index d7e3468..fd8b5e7 100644
--- a/src/microwink/seg.py
+++ b/src/microwink/seg.py
@@ -179,12 +179,12 @@ def postprocess(
         likely = x[:, 4 : 4 + NUM_CLASSES].max(axis=1) > conf_threshold
         x = x[likely]
 
-        boxes = x[:, :4]
         scores = x[:, 4 : 4 + NUM_CLASSES].max(axis=1)
+        boxes = x[:, :4]
+        boxes = self.postprocess_boxes(boxes, img_size, ratio, pad_w=pad_w, pad_h=pad_h)
         keep = self.nms(
             boxes,
             scores,
-            conf_threshold=conf_threshold,
             iou_threshold=iou_threshold,
         )
         N = len(keep)
@@ -199,7 +199,6 @@ def postprocess(
         masks_in = masks_in[keep]
 
         ih, iw = img_size
-        boxes = self.postprocess_boxes(boxes, img_size, ratio, pad_w=pad_w, pad_h=pad_h)
         masks = self.postprocess_masks(protos, masks_in, boxes, (ih, iw))
 
         assert masks.shape == (N, ih, iw)
@@ -293,21 +292,41 @@ def nms(
         boxes: np.ndarray,
         scores: np.ndarray,
         *,
-        conf_threshold: float,
         iou_threshold: float,
     ) -> list[int]:
-        from cv2.dnn import NMSBoxes
-
+        sorted_indices = np.argsort(scores)[::-1]
         N = len(boxes)
         assert boxes.shape == (N, 4)
         assert scores.shape == (N,)
-        keep = NMSBoxes(
-            boxes,  # type: ignore
-            scores,  # type: ignore
-            conf_threshold,
-            iou_threshold,
-        )
-        return list(keep)
+        assert sorted_indices.shape == (N,)
+
+        keep_boxes = []
+        while sorted_indices.size > 0:
+            box_id = int(sorted_indices[0])
+            ious = SegModel._compute_iou(boxes[box_id, :], boxes[sorted_indices[1:], :])
+            keep_indices = np.where(ious < iou_threshold)[0]
+            sorted_indices = sorted_indices[keep_indices + 1]
+
+            keep_boxes.append(box_id)
+        return keep_boxes
+
+    @staticmethod
+    def _compute_iou(box: np.ndarray, boxes: np.ndarray) -> np.ndarray:
+        assert box.shape == (4,)
+        assert boxes.shape == (len(boxes), 4)
+        xmin = np.maximum(box[0], boxes[:, 0])
+        ymin = np.maximum(box[1], boxes[:, 1])
+        xmax = np.minimum(box[2], boxes[:, 2])
+        ymax = np.minimum(box[3], boxes[:, 3])
+
+        intersection_area = np.maximum(0, xmax - xmin) * np.maximum(0, ymax - ymin)
+
+        box_area = (box[2] - box[0]) * (box[3] - box[1])
+        boxes_area = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
+        union_area = box_area + boxes_area - intersection_area
+
+        iou = intersection_area / union_area
+        return iou
 
     @staticmethod
     def with_border(
@@ -318,11 +337,11 @@ def with_border(
         right: int,
         color: tuple[int, int, int],
     ) -> np.ndarray:
-        from cv2 import BORDER_CONSTANT, copyMakeBorder
+        import cv2
 
         assert img.ndim == 3
-        return copyMakeBorder(
-            img, top, bottom, left, right, BORDER_CONSTANT, value=color
+        return cv2.copyMakeBorder(
+            img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color
         )