fix(workflows): translate oriented bounding box corners in dynamic_crop (#2430)

`crop_image` shifts every other detection geometry to crop-local coords
(xyxy via move_boxes, mask by slicing, keypoints and polygon by explicit
subtraction) but skipped `data['xyxyxyxy']`. OBB workflows that fed the
block (e.g. detect aerial vehicles, crop each one, run downstream OBB
analysis) ended up with corners still in the original image frame next
to a crop-local xyxy.

Same shape and pattern as the detections_stitch fix in #2427.

Author: kounelisagis

inference/core/workflows/core_steps/transformations/dynamic_crop/v1.py

@@ -5,6 +5,7 @@
 import numpy as np
 import supervision as sv
 from pydantic import AliasChoices, ConfigDict, Field
+from supervision.config import ORIENTED_BOX_COORDINATES
 
 from inference.core.workflows.execution_engine.constants import (
     DETECTION_ID_KEY,
@@ -256,6 +257,14 @@ def crop_image(
             translated_detection[POLYGON_KEY_IN_SV_DETECTIONS] = selected_detection[
                 POLYGON_KEY_IN_SV_DETECTIONS
             ] - np.array([x_min, y_min])
+        if ORIENTED_BOX_COORDINATES in detections.data:
+            # OBB corners (shape (1, 4, 2) on a single-detection slice) get the
+            # same crop-local offset applied to xyxy / keypoints / polygon. Without
+            # this, OBB-aware downstream blocks see corners in the original image
+            # frame next to a crop-local xyxy.
+            translated_detection[ORIENTED_BOX_COORDINATES] = selected_detection[
+                ORIENTED_BOX_COORDINATES
+            ] - np.array([x_min, y_min])
 
         crops.append(
             {

tests/workflows/unit_tests/core_steps/transformations/test_crop.py

@@ -5,6 +5,7 @@
 import pytest
 import supervision as sv
 from pydantic import ValidationError
+from supervision.config import ORIENTED_BOX_COORDINATES
 
 from inference.core.workflows.core_steps.transformations.dynamic_crop.v1 import (
     BlockManifest,
@@ -450,3 +451,68 @@ def test_crop_image_when_background_removal_requested_and_mask_found() -> None:
     assert (
         result[2]["crops"].numpy_image == expected_third_crop
     ).all(), "Image must have expected size and color"
+
+
+@pytest.mark.parametrize(
+    "crop_offset",
+    [
+        pytest.param((250, 300), id="interior-crop"),
+        pytest.param((0, 0), id="top-left-crop"),
+        pytest.param((100, 0), id="x-only-crop"),
+    ],
+)
+def test_crop_image_translates_oriented_bounding_box_corners(
+    crop_offset: Tuple[int, int],
+) -> None:
+    """OBB corners stored in `data['xyxyxyxy']` must follow the same `-(x_min, y_min)`
+    translation already applied to `xyxy`, `mask`, `keypoints`, and `polygon`.
+    Without this, downstream OBB-aware blocks would see corners in the original
+    image frame next to a crop-local `xyxy`."""
+    # given
+    x_min, y_min = crop_offset
+    side = 40
+    np_image = np.zeros((1000, 1000, 3), dtype=np.uint8)
+    image = WorkflowImageData(
+        parent_metadata=ImageParentMetadata(parent_id="origin_image"),
+        numpy_image=np_image,
+    )
+    # OBB centered inside the crop region, rotated 45-deg (diamond corners).
+    image_corners = np.array(
+        [
+            [
+                [x_min + side / 2, y_min],
+                [x_min + side, y_min + side / 2],
+                [x_min + side / 2, y_min + side],
+                [x_min, y_min + side / 2],
+            ]
+        ],
+        dtype=np.float32,
+    )
+    detections = sv.Detections(
+        xyxy=np.array([[x_min, y_min, x_min + side, y_min + side]], dtype=np.float64),
+        class_id=np.array([0]),
+        confidence=np.array([0.9], dtype=np.float64),
+        data={
+            "detection_id": np.array(["one"]),
+            "class_name": np.array(["thing"]),
+            ORIENTED_BOX_COORDINATES: image_corners,
+        },
+    )
+
+    # when
+    result = crop_image(
+        image=image,
+        detections=detections,
+        mask_opacity=0.0,
+        background_color=(0, 0, 0),
+    )
+
+    # then
+    assert len(result) == 1
+    translated = result[0]["predictions"]
+    expected_corners = image_corners - np.array([x_min, y_min], dtype=np.float32)
+    assert np.allclose(translated.data[ORIENTED_BOX_COORDINATES], expected_corners)
+    # Cross-check: AABB of the translated corners sits inside [0, side]^2.
+    translated_corners = translated.data[ORIENTED_BOX_COORDINATES][0]
+    assert translated_corners.min() >= 0.0
+    assert translated_corners.max() <= side