Add support for bounding boxes in croissant_builder.

PiperOrigin-RevId: 686902619

Author: The TensorFlow Datasets Authors

tensorflow_datasets/core/dataset_builders/croissant_builder.py

@@ -46,6 +46,7 @@
 from tensorflow_datasets.core import download
 from tensorflow_datasets.core import split_builder as split_builder_lib
 from tensorflow_datasets.core import splits as splits_lib
+from tensorflow_datasets.core.features import bounding_boxes
 from tensorflow_datasets.core.features import feature as feature_lib
 from tensorflow_datasets.core.features import features_dict
 from tensorflow_datasets.core.features import image_feature
@@ -86,12 +87,15 @@ def datatype_converter(
   if not field_data_type:
     # Fields with sub fields are of type None
     if field.sub_fields:
-      return features_dict.FeaturesDict({
-          subfield.id: datatype_converter(
-              subfield, int_dtype=int_dtype, float_dtype=float_dtype
-          )
-          for subfield in field.sub_fields
-      })
+      return features_dict.FeaturesDict(
+          {
+              subfield.id: datatype_converter(
+                  subfield, int_dtype=int_dtype, float_dtype=float_dtype
+              )
+              for subfield in field.sub_fields
+          },
+          doc=field.description,
+      )
     return None
   elif field_data_type == int:
     return int_dtype
@@ -106,6 +110,9 @@ def datatype_converter(
     return text_feature.Text(doc=field.description)
   elif field_data_type == mlc.DataType.IMAGE_OBJECT:
     return image_feature.Image(doc=field.description)
+  elif field_data_type == mlc.DataType.BOUNDING_BOX:
+    # TFDS uses REL_YXYX by default, but Hugging Face doesn't enforce a format.
+    return bounding_boxes.BBoxFeature(doc=field.description, bbox_format=None)
   else:
     raise ValueError(f'Unknown data type: {field_data_type}.')
 
@@ -225,7 +232,9 @@ def __init__(
   @property
   def builder_config(self) -> dataset_builder.BuilderConfig:
     """`tfds.core.BuilderConfig` for this builder."""
-    return self._builder_config  # pytype: disable=bad-return-type  # always-use-return-annotations
+    return (
+        self._builder_config
+    )  # pytype: disable=bad-return-type  # always-use-return-annotations
 
   def _info(self) -> dataset_info.DatasetInfo:
     return dataset_info.DatasetInfo(

tensorflow_datasets/core/dataset_builders/croissant_builder_test.py

@@ -20,6 +20,7 @@
 from tensorflow_datasets import testing
 from tensorflow_datasets.core import file_adapters
 from tensorflow_datasets.core.dataset_builders import croissant_builder
+from tensorflow_datasets.core.features import bounding_boxes
 from tensorflow_datasets.core.features import features_dict
 from tensorflow_datasets.core.features import image_feature
 from tensorflow_datasets.core.features import tensor_feature
@@ -117,6 +118,14 @@ def test_simple_datatype_converter(field, feature_type, int_dtype, float_dtype):
             image_feature.Image,
             None,
         ),
+        (
+            mlc.Field(
+                data_types=mlc.DataType.BOUNDING_BOX,
+                description="Bounding box feature",
+            ),
+            bounding_boxes.BBoxFeature,
+            None,
+        ),
         (
             mlc.Field(
                 id="person",
@@ -138,6 +147,7 @@ def test_simple_datatype_converter(field, feature_type, int_dtype, float_dtype):
 def test_complex_datatype_converter(field, feature_type, subfield_types):
   actual_feature = croissant_builder.datatype_converter(field)
   assert isinstance(actual_feature, feature_type)
+  assert actual_feature.doc.desc == field.description
   if subfield_types:
     for feature_name in actual_feature.keys():
       assert isinstance(