Add attributes in VOC format (#1792)

* Add voc attributes

* Allow any values for voc pose

* update changelog

* Add attribute conversion

* linter

* fix tests

CHANGELOG.md

@@ -32,6 +32,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Settings page move to the modal. (<https://github.com/opencv/cvat/pull/1705>)
 - Implemented import and export of annotations with relative image paths (<https://github.com/opencv/cvat/pull/1463>)
 - Using only single click to start editing or remove a point (<https://github.com/opencv/cvat/pull/1571>)
+- Added support for attributes in VOC XML format (https://github.com/opencv/cvat/pull/1792)
 
 ### Deprecated
 -

cvat/apps/dataset_manager/bindings.py

@@ -49,7 +49,8 @@ class TaskData:
             (db_label.id, db_label) for db_label in db_labels)
 
         self._attribute_mapping = {db_label.id: {
-            'mutable': {}, 'immutable': {}} for db_label in db_labels}
+            'mutable': {}, 'immutable': {}, 'spec': {}}
+            for db_label in db_labels}
 
         for db_label in db_labels:
             for db_attribute in db_label.attributespec_set.all():
@@ -57,6 +58,7 @@ class TaskData:
                     self._attribute_mapping[db_label.id]['mutable'][db_attribute.id] = db_attribute.name
                 else:
                     self._attribute_mapping[db_label.id]['immutable'][db_attribute.id] = db_attribute.name
+                self._attribute_mapping[db_label.id]['spec'][db_attribute.id] = db_attribute
 
         self._attribute_mapping_merged = {}
         for label_id, attr_mapping in self._attribute_mapping.items():
@@ -317,10 +319,28 @@ class TaskData:
         return _tag
 
     def _import_attribute(self, label_id, attribute):
-        return {
-            'spec_id': self._get_attribute_id(label_id, attribute.name),
-            'value': attribute.value,
-        }
+        spec_id = self._get_attribute_id(label_id, attribute.name)
+        value = attribute.value
+
+        if spec_id:
+            spec = self._attribute_mapping[label_id]['spec'][spec_id]
+
+            try:
+                if spec.input_type == AttributeType.NUMBER:
+                    pass # no extra processing required
+                elif spec.input_type == AttributeType.CHECKBOX:
+                    if isinstance(value, str):
+                        value = value.lower()
+                        assert value in {'true', 'false'}
+                    elif isinstance(value, (bool, int, float)):
+                        value = 'true' if value else 'false'
+                    else:
+                        raise ValueError("Unexpected attribute value")
+            except Exception as e:
+                raise Exception("Failed to convert attribute '%s'='%s': %s" %
+                    (self._get_label_name(label_id), value, e))
+
+        return { 'spec_id': spec_id, 'value': value }
 
     def _import_shape(self, shape):
         _shape = shape._asdict()

cvat/apps/engine/tests/_test_rest_api.py

@@ -1988,7 +1988,7 @@ class JobAnnotationAPITestCase(APITestCase):
                             "name": "parked",
                             "mutable": True,
                             "input_type": "checkbox",
-                            "default_value": False
+                            "default_value": "false"
                         },
                     ]
                 },

datumaro/datumaro/plugins/voc_format/converter.py

@@ -53,7 +53,8 @@ LabelmapType = Enum('LabelmapType', ['voc', 'source', 'guess'])
 
 class _Converter:
     def __init__(self, extractor, save_dir,
-            tasks=None, apply_colormap=True, save_images=False, label_map=None):
+            tasks=None, apply_colormap=True, save_images=False, label_map=None,
+            allow_attributes=True):
         assert tasks is None or isinstance(tasks, (VocTask, list, set))
         if tasks is None:
             tasks = set(VocTask)
@@ -66,6 +67,7 @@ class _Converter:
         self._extractor = extractor
         self._save_dir = save_dir
         self._apply_colormap = apply_colormap
+        self._allow_attributes = allow_attributes
         self._save_images = save_images
 
         self._load_categories(label_map)
@@ -205,9 +207,8 @@ class _Converter:
                         ET.SubElement(obj_elem, 'name').text = obj_label
 
                         if 'pose' in attr:
-                            pose = _convert_attr('pose', attr,
-                                lambda v: VocPose[v], VocPose.Unspecified)
-                            ET.SubElement(obj_elem, 'pose').text = pose.name
+                            ET.SubElement(obj_elem, 'pose').text = \
+                                str(attr['pose'])
 
                         if 'truncated' in attr:
                             truncated = _convert_attr('truncated', attr, int, 0)
@@ -252,6 +253,21 @@ class _Converter:
                         if len(actions_elem) != 0:
                             obj_elem.append(actions_elem)
 
+                        if self._allow_attributes:
+                            native_attrs = {'difficult', 'pose',
+                                'truncated', 'occluded' }
+                            native_attrs.update(label_actions)
+
+                            attrs_elem = ET.Element('attributes')
+                            for k, v in attr.items():
+                                if k in native_attrs:
+                                    continue
+                                attr_elem = ET.SubElement(attrs_elem, 'attribute')
+                                ET.SubElement(attr_elem, 'name').text = str(k)
+                                ET.SubElement(attr_elem, 'value').text = str(v)
+                            if len(attrs_elem):
+                                obj_elem.append(attrs_elem)
+
                     if self._tasks & {None,
                             VocTask.detection,
                             VocTask.person_layout,
@@ -565,15 +581,17 @@ class VocConverter(Converter, CliPlugin):
         parser.add_argument('--label-map', type=cls._get_labelmap, default=None,
             help="Labelmap file path or one of %s" % \
                 ', '.join(t.name for t in LabelmapType))
+        parser.add_argument('--allow-attributes',
+            type=str_to_bool, default=True,
+            help="Allow export of attributes (default: %(default)s)")
         parser.add_argument('--tasks', type=cls._split_tasks_string,
-            default=None,
             help="VOC task filter, comma-separated list of {%s} "
-                "(default: all)" % ', '.join([t.name for t in VocTask]))
+                "(default: all)" % ', '.join(t.name for t in VocTask))
 
         return parser
 
     def __init__(self, tasks=None, save_images=False,
-            apply_colormap=False, label_map=None):
+            apply_colormap=False, label_map=None, allow_attributes=True):
         super().__init__()
 
         self._options = {
@@ -581,6 +599,7 @@ class VocConverter(Converter, CliPlugin):
             'save_images': save_images,
             'apply_colormap': apply_colormap,
             'label_map': label_map,
+            'allow_attributes': allow_attributes,
         }
 
     def __call__(self, extractor, save_dir):

datumaro/datumaro/plugins/voc_format/extractor.py

@@ -198,6 +198,12 @@ class _VocXmlExtractor(_VocExtractor):
                 item_annotations.append(Bbox(*part_bbox, label=part_label_id,
                     group=group))
 
+            attributes_elem = object_elem.find('attributes')
+            if attributes_elem is not None:
+                for attr_elem in attributes_elem.iter('attribute'):
+                    attributes[attr_elem.find('name').text] = \
+                        attr_elem.find('value').text
+
             if self._task is VocTask.person_layout and not has_parts:
                 continue
             if self._task is VocTask.action_classification and not actions:

datumaro/tests/test_voc_format.py

@@ -395,8 +395,8 @@ class VocConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                VocActionConverter(label_map='voc'), test_dir,
-                target_dataset=DstExtractor())
+                VocActionConverter(label_map='voc', allow_attributes=False),
+                test_dir, target_dataset=DstExtractor())
 
     def test_can_save_dataset_with_no_subsets(self):
         class TestExtractor(TestExtractorBase):
@@ -679,3 +679,34 @@ class VocConverterTest(TestCase):
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
                 VocConverter(label_map='voc', save_images=True), test_dir)
+
+    def test_can_save_attributes(self):
+        class TestExtractor(TestExtractorBase):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='a', annotations=[
+                        Bbox(2, 3, 4, 5, label=2,
+                            attributes={ 'occluded': True, 'x': 1, 'y': '2' }
+                        ),
+                    ]),
+                ])
+
+        class DstExtractor(TestExtractorBase):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='a', annotations=[
+                        Bbox(2, 3, 4, 5, label=2, id=1, group=1,
+                            attributes={
+                                'truncated': False,
+                                'difficult': False,
+                                'occluded': True,
+                                'x': '1', 'y': '2', # can only read strings
+                            }
+                        ),
+                    ]),
+                ])
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(TestExtractor(),
+                VocDetectionConverter(label_map='voc'), test_dir,
+                target_dataset=DstExtractor())