Fix box filtering

albumentations/augmentations/crops/transforms.py

@@ -20,6 +20,7 @@
     ZeroOneRangeType,
     check_0plus,
     check_01,
+    nondecreasing,
 )
 from albumentations.core.transforms_interface import BaseTransformInitSchema, DualTransform
 from albumentations.core.types import (
@@ -672,8 +673,8 @@ class RandomResizedCrop(_BaseRandomSizedCrop):
     _targets = (Targets.IMAGE, Targets.MASK, Targets.BBOXES, Targets.KEYPOINTS)
 
     class InitSchema(BaseTransformInitSchema):
-        scale: Annotated[tuple[float, float], AfterValidator(check_01)]
-        ratio: Annotated[tuple[float, float], AfterValidator(check_0plus)]
+        scale: Annotated[tuple[float, float], AfterValidator(check_01), AfterValidator(nondecreasing)]
+        ratio: Annotated[tuple[float, float], AfterValidator(check_0plus), AfterValidator(nondecreasing)]
         width: int | None = Field(
             None,
             deprecated="Initializing with 'height' and 'width' is deprecated. Use size instead.",

albumentations/core/utils.py

@@ -132,16 +132,28 @@ def ensure_transforms_valid(self, transforms: Sequence[object]) -> None:
 
     def postprocess(self, data: dict[str, Any]) -> dict[str, Any]:
         image_shape = get_shape(data["image"])
+        data = self._process_data_fields(data, image_shape)
         data = self.remove_label_fields_from_data(data)
+        return self._convert_sequence_inputs(data)
 
+    def _process_data_fields(self, data: dict[str, Any], image_shape: tuple[int, int]) -> dict[str, Any]:
         for data_name in set(self.data_fields) & set(data.keys()):
-            data[data_name] = self.filter(data[data_name], image_shape)
+            data[data_name] = self._process_single_field(data_name, data[data_name], image_shape)
+        return data
+
+    def _process_single_field(self, data_name: str, field_data: Any, image_shape: tuple[int, int]) -> Any:
+        field_data = self.filter(field_data, image_shape)
+
+        if data_name == "keypoints" and len(field_data) == 0:
+            field_data = self._create_empty_keypoints_array()
 
-            if data_name == "keypoints" and len(data[data_name]) == 0:
-                data[data_name] = np.array([], dtype=np.float32).reshape(0, len(self.params.format))
+        return self.check_and_convert(field_data, image_shape, direction="from")
 
-            data[data_name] = self.check_and_convert(data[data_name], image_shape, direction="from")
-            # Convert back to list of lists if original input was a list
+    def _create_empty_keypoints_array(self) -> np.ndarray:
+        return np.array([], dtype=np.float32).reshape(0, len(self.params.format))
+
+    def _convert_sequence_inputs(self, data: dict[str, Any]) -> dict[str, Any]:
+        for data_name in set(self.data_fields) & set(data.keys()):
             if self.is_sequence_input.get(data_name, False):
                 data[data_name] = data[data_name].tolist()
         return data

tests/test_bbox.py

@@ -1522,3 +1522,19 @@ def test_bboxes_from_masks_output_type():
     result = bboxes_from_masks(masks)
     assert isinstance(result, np.ndarray)
     assert result.dtype == np.int32
+
+
+def test_random_resized_crop():
+    transform = A.Compose(
+    [
+        A.RandomResizedCrop((100, 100), scale=(0.01, 0.1), ratio=(1, 1)),
+    ],
+    bbox_params=A.BboxParams(
+        format="coco",
+        label_fields=["label"],
+    ),
+    )
+    boxes = [[10,10,20,20], [5,5,10,10], [450, 450, 5,5], [250,250,5,5]]
+    labels = [1,2,3,4]
+    res = transform(image=np.zeros((500,500,3), dtype='uint8'), bboxes=boxes, label=labels)
+    assert len(res['bboxes']) == len(res['label'])