Merge pull request #956 from roboflow/detection-offset-percent

Added Percent Padding to the Detection Offset Block

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

@@ -75,6 +75,11 @@ class BlockManifest(WorkflowBlockManifest):
         examples=[10, "$inputs.offset_y"],
         validation_alias=AliasChoices("offset_height", "offset_y"),
     )
+    units: Literal["Percent (%)", "Pixels"] = Field(
+        default="Pixels",
+        description="Units for offset dimensions",
+        examples=["Pixels", "Percent (%)"],
+    )
 
     @classmethod
     def get_parameters_accepting_batches(cls) -> List[str]:
@@ -111,13 +116,16 @@ def run(
         predictions: Batch[sv.Detections],
         offset_width: int,
         offset_height: int,
+        units: str = "Pixels",
     ) -> BlockResult:
+        use_percentage = units == "Percent (%)"
         return [
             {
                 "predictions": offset_detections(
                     detections=detections,
                     offset_width=offset_width,
                     offset_height=offset_height,
+                    use_percentage=use_percentage,
                 )
             }
             for detections in predictions
@@ -130,22 +138,42 @@ def offset_detections(
     offset_height: int,
     parent_id_key: str = PARENT_ID_KEY,
     detection_id_key: str = DETECTION_ID_KEY,
+    use_percentage: bool = False,
 ) -> sv.Detections:
     if len(detections) == 0:
         return detections
     _detections = deepcopy(detections)
     image_dimensions = detections.data["image_dimensions"]
-    _detections.xyxy = np.array(
-        [
-            (
-                max(0, x1 - offset_width // 2),
-                max(0, y1 - offset_height // 2),
-                min(image_dimensions[i][1], x2 + offset_width // 2),
-                min(image_dimensions[i][0], y2 + offset_height // 2),
-            )
-            for i, (x1, y1, x2, y2) in enumerate(_detections.xyxy)
-        ]
-    )
+    if use_percentage:
+        _detections.xyxy = np.array(
+            [
+                (
+                    max(0, x1 - int(image_dimensions[i][1] * offset_width / 200)),
+                    max(0, y1 - int(image_dimensions[i][0] * offset_height / 200)),
+                    min(
+                        image_dimensions[i][1],
+                        x2 + int(image_dimensions[i][1] * offset_width / 200),
+                    ),
+                    min(
+                        image_dimensions[i][0],
+                        y2 + int(image_dimensions[i][0] * offset_height / 200),
+                    ),
+                )
+                for i, (x1, y1, x2, y2) in enumerate(_detections.xyxy)
+            ]
+        )
+    else:
+        _detections.xyxy = np.array(
+            [
+                (
+                    max(0, x1 - offset_width // 2),
+                    max(0, y1 - offset_height // 2),
+                    min(image_dimensions[i][1], x2 + offset_width // 2),
+                    min(image_dimensions[i][0], y2 + offset_height // 2),
+                )
+                for i, (x1, y1, x2, y2) in enumerate(_detections.xyxy)
+            ]
+        )
     _detections[parent_id_key] = detections[detection_id_key].copy()
     _detections[detection_id_key] = [str(uuid.uuid4()) for _ in detections]
     return _detections

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

@@ -158,3 +158,33 @@ def test_offset_detection_when_nothing_predicted() -> None:
 
     # then
     assert len(detections) == 0, "Expected empty detections in output"
+
+
+def test_offset_detection_with_percentage() -> None:
+    # given
+    detections = sv.Detections(
+        xyxy=np.array([[100, 200, 300, 400]], dtype=np.float64),
+        class_id=np.array([1]),
+        confidence=np.array([0.5], dtype=np.float64),
+        data={
+            "detection_id": np.array(["three"]),
+            "class_name": np.array(["truck"]),
+            "parent_id": np.array(["p3"]),
+            "image_dimensions": np.array([[640, 640]]),
+        },
+    )
+
+    # when
+    result = offset_detections(
+        detections=detections,
+        offset_width=10,
+        offset_height=10,
+        use_percentage=True
+    )
+
+    # then
+    x1, y1, x2, y2 = result.xyxy[0]
+    assert x1 == 68, "Left corner should be moved by 10% of image width to the left"
+    assert y1 == 168, "Top corner should be moved by 10% of image height to the top"
+    assert x2 == 332, "Right corner should be moved by 10% of image width to the right"
+    assert y2 == 432, "Bottom corner should be moved by 10% of image height to the bottom"