feat: ✨ scale_detections function added to adjust

    bounding box coordinates and masks and obb for scaled images

Signed-off-by: Onuralp SEZER <[email protected]>

supervision/__init__.py

@@ -40,6 +40,7 @@
 )
 from supervision.dataset.utils import mask_to_rle, rle_to_mask
 from supervision.detection.core import Detections
+from supervision.detection.detection_utils import scale_detections
 from supervision.detection.line_zone import (
     LineZone,
     LineZoneAnnotator,
@@ -219,6 +220,7 @@
     "resize_image",
     "rle_to_mask",
     "scale_boxes",
+    "scale_detections",
     "scale_image",
     "xcycwh_to_xyxy",
     "xywh_to_xyxy",

supervision/detection/detection_utils.py

@@ -0,0 +1,82 @@
+from typing import Tuple
+
+import cv2
+import numpy as np
+
+from supervision.detection.core import ORIENTED_BOX_COORDINATES, Detections
+
+
+def scale_detections(
+    detections: Detections,
+    letterbox_wh: Tuple[int, int],
+    resolution_wh: Tuple[int, int],
+) -> Detections:
+    """
+    This function scale the coordinates of bounding boxes and optionally scales the
+    masks,oriented bounding boxes to fit a new resolution, taking into account the
+    letterbox padding applied during the resizing process and return Detections object.
+
+    Args:
+        detections (Detections): The Detections object to be scaled.
+        letterbox_wh (Tuple[int, int]): The width and height of the letterboxed image.
+        resolution_wh (Tuple[int, int]): The target width and height for scaling.
+
+    Returns:
+        Detections: A new Detections object with scaled to target resolution.
+    """
+    input_w, input_h = resolution_wh
+    letterbox_w, letterbox_h = letterbox_wh
+
+    target_ratio = letterbox_w / letterbox_h
+    image_ratio = input_w / input_h
+
+    if image_ratio >= target_ratio:
+        width_new = letterbox_w
+        height_new = int(letterbox_w / image_ratio)
+    else:
+        height_new = letterbox_h
+        width_new = int(letterbox_h * image_ratio)
+
+    scale = input_w / width_new
+    padding_top = (letterbox_h - height_new) // 2
+    padding_left = (letterbox_w - width_new) // 2
+
+    boxes = detections.xyxy.copy()
+    boxes[:, [0, 2]] -= padding_left
+    boxes[:, [1, 3]] -= padding_top
+    boxes[:, [0, 2]] *= scale
+    boxes[:, [1, 3]] *= scale
+
+    scaled_mask = None
+    if detections.mask is not None:
+        masks = []
+        for mask in detections.mask:
+            mask = mask[
+                padding_top : padding_top + height_new,
+                padding_left : padding_left + width_new,
+            ]
+            scaled_mask_i = cv2.resize(
+                mask.astype(np.uint8),
+                (input_w, input_h),
+                interpolation=cv2.INTER_LINEAR,
+            ).astype(bool)
+            masks.append(scaled_mask_i)
+        scaled_mask = np.array(masks)
+
+    if ORIENTED_BOX_COORDINATES in detections.data:
+        obbs = np.array(detections.data[ORIENTED_BOX_COORDINATES]).copy()
+        obbs[:, :, 0] -= padding_left
+        obbs[:, :, 1] -= padding_top
+        obbs[:, :, 0] *= scale
+        obbs[:, :, 1] *= scale
+        detections.data[ORIENTED_BOX_COORDINATES] = obbs
+
+    return Detections(
+        xyxy=boxes,
+        mask=scaled_mask,
+        confidence=detections.confidence,
+        class_id=detections.class_id,
+        tracker_id=detections.tracker_id,
+        data=detections.data,
+        metadata=detections.metadata,
+    )