[Datumaro] Support relative paths (#1715)

* Support relative image paths in Datumaro

* Update bindings

* Fix merge

* linter

cvat/apps/dataset_manager/bindings.py

@@ -11,6 +11,7 @@ from django.utils import timezone
 import datumaro.components.extractor as datumaro
 from cvat.apps.engine.frame_provider import FrameProvider
 from cvat.apps.engine.models import AttributeType, ShapeType
+from datumaro.util import cast
 from datumaro.util.image import Image
 
 from .annotation import AnnotationManager, TrackManager
@@ -422,8 +423,9 @@ class CvatTaskDataExtractor(datumaro.SourceExtractor):
                 size=(frame_data.height, frame_data.width)
             )
             dm_anno = self._read_cvat_anno(frame_data, task_data)
-            dm_item = datumaro.DatasetItem(id=frame_data.frame,
-                annotations=dm_anno, image=dm_image)
+            dm_item = datumaro.DatasetItem(id=osp.splitext(frame_data.name)[0],
+                annotations=dm_anno, image=dm_image,
+                attributes={'frame': frame_data.frame})
             dm_items.append(dm_item)
 
         self._items = dm_items
@@ -533,23 +535,21 @@ def match_frame(item, task_data):
     is_video = task_data.meta['task']['mode'] == 'interpolation'
 
     frame_number = None
-    if frame_number is None:
-        try:
-            frame_number = task_data.match_frame(item.id)
-        except Exception:
-            pass
     if frame_number is None and item.has_image:
         try:
-            frame_number = task_data.match_frame(item.image.filename)
+            frame_number = task_data.match_frame(item.image.path)
         except Exception:
             pass
     if frame_number is None:
         try:
-            frame_number = int(item.id)
+            frame_number = task_data.match_frame(item.id)
         except Exception:
             pass
-    if frame_number is None and is_video and item.id.startswith('frame_'):
-        frame_number = int(item.id[len('frame_'):])
+    if frame_number is None:
+        frame_number = cast(item.attributes.get('frame', item.id), int)
+    if frame_number is None and is_video:
+        frame_number = cast(osp.basename(item.id)[len('frame_'):], int)
+
     if not frame_number in task_data.frame_info:
         raise Exception("Could not match item id: '%s' with any task frame" %
             item.id)

cvat/apps/dataset_manager/formats/labelme.py

@@ -17,8 +17,6 @@ from .registry import dm_env, exporter, importer
 @exporter(name='LabelMe', ext='ZIP', version='3.0')
 def _export(dst_file, task_data, save_images=False):
     extractor = CvatTaskDataExtractor(task_data, include_images=save_images)
-    envt = dm_env.transforms
-    extractor = extractor.transform(envt.get('id_from_image_name'))
     extractor = Dataset.from_extractors(extractor) # apply lazy transforms
     with TemporaryDirectory() as temp_dir:
         converter = dm_env.make_converter('label_me', save_images=save_images)

cvat/apps/dataset_manager/formats/mask.py

@@ -24,7 +24,6 @@ def _export(dst_file, task_data, save_images=False):
     extractor = extractor.transform(envt.get('polygons_to_masks'))
     extractor = extractor.transform(envt.get('boxes_to_masks'))
     extractor = extractor.transform(envt.get('merge_instance_segments'))
-    extractor = extractor.transform(envt.get('id_from_image_name'))
     extractor = Dataset.from_extractors(extractor) # apply lazy transforms
     with TemporaryDirectory() as temp_dir:
         converter = dm_env.make_converter('voc_segmentation',

cvat/apps/dataset_manager/formats/mot.py

@@ -7,8 +7,7 @@ from tempfile import TemporaryDirectory
 from pyunpack import Archive
 
 import datumaro.components.extractor as datumaro
-from cvat.apps.dataset_manager.bindings import (CvatTaskDataExtractor,
-    match_frame)
+from cvat.apps.dataset_manager.bindings import CvatTaskDataExtractor
 from cvat.apps.dataset_manager.util import make_zip_archive
 from datumaro.components.project import Dataset
 
@@ -18,8 +17,6 @@ from .registry import dm_env, exporter, importer
 @exporter(name='MOT', ext='ZIP', version='1.1')
 def _export(dst_file, task_data, save_images=False):
     extractor = CvatTaskDataExtractor(task_data, include_images=save_images)
-    envt = dm_env.transforms
-    extractor = extractor.transform(envt.get('id_from_image_name'))
     extractor = Dataset.from_extractors(extractor) # apply lazy transforms
     with TemporaryDirectory() as temp_dir:
         converter = dm_env.make_converter('mot_seq_gt',
@@ -39,8 +36,8 @@ def _import(src_file, task_data):
         label_cat = dataset.categories()[datumaro.AnnotationType.label]
 
         for item in dataset:
-            item = item.wrap(id=int(item.id) - 1) # NOTE: MOT frames start from 1
-            frame_number = task_data.abs_frame_id(match_frame(item, task_data))
+            frame_number = int(item.id) - 1 # NOTE: MOT frames start from 1
+            frame_number = task_data.abs_frame_id(frame_number)
 
             for ann in item.annotations:
                 if ann.type != datumaro.AnnotationType.bbox:

cvat/apps/dataset_manager/formats/pascal_voc.py

@@ -22,8 +22,6 @@ from .registry import dm_env, exporter, importer
 @exporter(name='PASCAL VOC', ext='ZIP', version='1.1')
 def _export(dst_file, task_data, save_images=False):
     extractor = CvatTaskDataExtractor(task_data, include_images=save_images)
-    envt = dm_env.transforms
-    extractor = extractor.transform(envt.get('id_from_image_name'))
     extractor = Dataset.from_extractors(extractor) # apply lazy transforms
     with TemporaryDirectory() as temp_dir:
         converter = dm_env.make_converter('voc', label_map='source',

datumaro/datumaro/components/extractor.py

@@ -578,7 +578,7 @@ class DatasetItem:
     def __init__(self, id=None, annotations=None,
             subset=None, path=None, image=None, attributes=None):
         assert id is not None
-        self._id = str(id)
+        self._id = str(id).replace('\\', '/')
 
         if subset is None:
             subset = ''

datumaro/datumaro/plugins/coco_format/converter.py

@@ -496,25 +496,19 @@ class _Converter:
     def _get_image_id(self, item):
         image_id = self._image_ids.get(item.id)
         if image_id is None:
-            image_id = cast(item.id, int, len(self._image_ids) + 1)
+            image_id = cast(item.attributes.get('id'), int,
+                len(self._image_ids) + 1)
             self._image_ids[item.id] = image_id
         return image_id
 
-    def _save_image(self, item):
+    def _save_image(self, item, filename):
         image = item.image.data
         if image is None:
             log.warning("Item '%s' has no image" % item.id)
             return ''
 
-        filename = item.image.filename
-        if filename:
-            filename = osp.splitext(filename)[0]
-        else:
-            filename = item.id
-        filename += CocoPath.IMAGE_EXT
-        path = osp.join(self._images_dir, filename)
-        save_image(path, image)
-        return path
+        save_image(osp.join(self._images_dir, filename), image,
+            create_dir=True)
 
     def convert(self):
         self._make_dirs()
@@ -534,12 +528,10 @@ class _Converter:
             for task_conv in task_converters.values():
                 task_conv.save_categories(subset)
             for item in subset:
-                filename = ''
-                if item.has_image:
-                    filename = item.image.path
+                filename = item.id + CocoPath.IMAGE_EXT
                 if self._save_images:
                     if item.has_image:
-                        filename = self._save_image(item)
+                        self._save_image(item, filename)
                     else:
                         log.debug("Item '%s' has no image info" % item.id)
                 for task_conv in task_converters.values():

datumaro/datumaro/plugins/coco_format/extractor.py

@@ -126,8 +126,10 @@ class _CocoExtractor(SourceExtractor):
             anns = loader.loadAnns(anns)
             anns = sum((self._load_annotations(a, image_info) for a in anns), [])
 
-            items[img_id] = DatasetItem(id=img_id, subset=self._subset,
-                image=image, annotations=anns)
+            items[img_id] = DatasetItem(
+                id=osp.splitext(image_info['file_name'])[0],
+                subset=self._subset, image=image, annotations=anns,
+                attributes={'id': img_id})
 
         return items
 

datumaro/datumaro/plugins/cvat_format/converter.py

@@ -163,26 +163,21 @@ class _SubsetWriter:
 
         self._writer.close_root()
 
-    def _save_image(self, item):
+    def _save_image(self, item, filename):
         image = item.image.data
         if image is None:
             log.warning("Item '%s' has no image" % item.id)
             return ''
 
-        filename = item.image.filename
-        if filename:
-            filename = osp.splitext(filename)[0]
-        else:
-            filename = item.id
-        filename += CvatPath.IMAGE_EXT
-        image_path = osp.join(self._context._images_dir, filename)
-        save_image(image_path, image)
-        return filename
+        save_image(osp.join(self._context._images_dir, filename), image,
+            create_dir=True)
 
     def _write_item(self, item, index):
         image_info = OrderedDict([
-            ("id", str(cast(item.id, int, index))),
+            ("id", str(cast(item.attributes.get('frame'), int, index))),
         ])
+        filename = item.id + CvatPath.IMAGE_EXT
+        image_info["name"] = filename
         if item.has_image:
             size = item.image.size
             if size:
@@ -190,10 +185,8 @@ class _SubsetWriter:
                 image_info["width"] = str(w)
                 image_info["height"] = str(h)
 
-            filename = item.image.filename
             if self._context._save_images:
-                filename = self._save_image(item)
-            image_info["name"] = filename
+                self._save_image(item, filename)
         else:
             log.debug("Item '%s' has no image info" % item.id)
         self._writer.open_image(image_info)

datumaro/datumaro/plugins/cvat_format/extractor.py

@@ -303,17 +303,14 @@ class CvatExtractor(SourceExtractor):
 
     def _load_items(self, parsed):
         for frame_id, item_desc in parsed.items():
-            path = item_desc.get('name', 'frame_%06d.png' % int(frame_id))
+            name = item_desc.get('name', 'frame_%06d.png' % int(frame_id))
+            image = osp.join(self._images_dir, name)
             image_size = (item_desc.get('height'), item_desc.get('width'))
             if all(image_size):
-                image_size = (int(image_size[0]), int(image_size[1]))
-            else:
-                image_size = None
-            image = None
-            if path:
-                image = Image(path=osp.join(self._images_dir, path),
-                    size=image_size)
-
-            parsed[frame_id] = DatasetItem(id=frame_id, subset=self._subset,
-                image=image, annotations=item_desc.get('annotations'))
+                image = Image(path=image, size=tuple(map(int, image_size)))
+
+            parsed[frame_id] = DatasetItem(id=osp.splitext(name)[0],
+                subset=self._subset, image=image,
+                annotations=item_desc.get('annotations'),
+                attributes={'frame': int(frame_id)})
         return parsed

datumaro/datumaro/plugins/datumaro_format/converter.py

@@ -253,12 +253,7 @@ class _Converter:
         if image is None:
             return ''
 
-        filename = item.image.filename
-        if filename:
-            filename = osp.splitext(filename)[0]
-        else:
-            filename = item.id
-        filename += DatumaroPath.IMAGE_EXT
+        filename = item.id + DatumaroPath.IMAGE_EXT
         image_path = osp.join(self._images_dir, filename)
         save_image(image_path, image, create_dir=True)
         return filename

datumaro/datumaro/plugins/datumaro_format/extractor.py

@@ -82,10 +82,11 @@ class DatumaroExtractor(SourceExtractor):
             item_id = item_desc['id']
 
             image = None
-            image_info = item_desc.get('image', {})
+            image_info = item_desc.get('image')
             if image_info:
-                image_path = osp.join(self._images_dir,
-                    image_info.get('path', '')) # relative or absolute fits
+                image_path = image_info.get('path') or \
+                    item_id + DatumaroPath.IMAGE_EXT
+                image_path = osp.join(self._images_dir, image_path)
                 image = Image(path=image_path, size=image_info.get('size'))
 
             annotations = self._load_annotations(item_desc)

datumaro/datumaro/plugins/image_dir.py

@@ -3,7 +3,6 @@
 #
 # SPDX-License-Identifier: MIT
 
-from collections import OrderedDict
 import os
 import os.path as osp
 
@@ -41,29 +40,24 @@ class ImageDirExtractor(SourceExtractor):
         assert osp.isdir(url), url
 
         items = []
-        for name in os.listdir(url):
-            path = osp.join(url, name)
-            if self._is_image(path):
-                item_id = osp.splitext(name)[0]
-                item = DatasetItem(id=item_id, image=path)
-                items.append((item.id, item))
-
-        items = sorted(items, key=lambda e: e[0])
-        items = OrderedDict(items)
+        for dirpath, _, filenames in os.walk(url):
+            for name in filenames:
+                path = osp.join(dirpath, name)
+                if not self._is_image(path):
+                    continue
+
+                item_id = osp.relpath(osp.splitext(path)[0], url)
+                items.append(DatasetItem(id=item_id, image=path))
+
         self._items = items
 
     def __iter__(self):
-        for item in self._items.values():
+        for item in self._items:
             yield item
 
     def __len__(self):
         return len(self._items)
 
-    def get(self, item_id, subset=None, path=None):
-        if path or subset:
-            raise KeyError()
-        return self._items[item_id]
-
     def _is_image(self, path):
         if not osp.isfile(path):
             return False
@@ -79,11 +73,5 @@ class ImageDirConverter(Converter):
 
         for item in extractor:
             if item.has_image and item.image.has_data:
-                filename = item.image.filename
-                if filename:
-                    filename = osp.splitext(filename)[0]
-                else:
-                    filename = item.id
-                filename += '.jpg'
-                save_image(osp.join(save_dir, filename), item.image.data,
-                    create_dir=True)
+                save_image(osp.join(save_dir, item.id + '.jpg'),
+                    item.image.data, create_dir=True)

datumaro/datumaro/plugins/labelme_format.py

@@ -331,16 +331,13 @@ class LabelMeConverter(Converter, CliPlugin):
 
         log.debug("Converting item '%s'", item.id)
 
-        image_filename = ''
-        if item.has_image:
-            image_filename = item.image.filename
+        if '/' in item.id:
+            raise Exception("Can't export item '%s': "
+                "LabelMe format only supports flat image layout" % item.id)
+
+        image_filename = item.id + LabelMePath.IMAGE_EXT
         if self._save_images:
             if item.has_image and item.image.has_data:
-                if image_filename:
-                    image_filename = osp.splitext(image_filename)[0]
-                else:
-                    image_filename = item.id
-                image_filename += LabelMePath.IMAGE_EXT
                 save_image(osp.join(subset_dir, image_filename),
                     item.image.data, create_dir=True)
             else:

datumaro/datumaro/plugins/mot_format.py

@@ -311,7 +311,9 @@ class MotSeqGtConverter(Converter, CliPlugin):
 
                 if self._save_images:
                     if item.has_image and item.image.has_data:
-                        self._save_image(item, index=frame_id)
+                        save_image(osp.join(self._images_dir,
+                                '%06d%s' % (frame_id, MotPath.IMAGE_EXT)),
+                            item.image.data)
                     else:
                         log.debug("Item '%s' has no image" % item.id)
 
@@ -320,13 +322,3 @@ class MotSeqGtConverter(Converter, CliPlugin):
             f.write('\n'.join(l.name
                 for l in extractor.categories()[AnnotationType.label].items)
             )
-
-    def _save_image(self, item, index):
-        if item.image.filename:
-            frame_id = osp.splitext(item.image.filename)[0]
-        else:
-            frame_id = item.id
-        frame_id = cast(frame_id, int, index)
-        image_filename = '%06d%s' % (frame_id, MotPath.IMAGE_EXT)
-        save_image(osp.join(self._images_dir, image_filename),
-            item.image.data)

datumaro/datumaro/plugins/tf_detection_api_format/converter.py

@@ -162,14 +162,12 @@ class TfDetectionApiConverter(Converter, CliPlugin):
 
     def _make_tf_example(self, item):
         features = {
-            'image/source_id': bytes_feature(str(item.id).encode('utf-8')),
+            'image/source_id': bytes_feature(
+                str(item.attributes.get('source_id') or '').encode('utf-8')
+            ),
         }
 
-        filename = ''
-        if item.has_image:
-            filename = item.image.filename
-        if not filename:
-            filename = item.id + DetectionApiPath.IMAGE_EXT
+        filename = item.id + DetectionApiPath.IMAGE_EXT
         features['image/filename'] = bytes_feature(filename.encode('utf-8'))
 
         if not item.has_image:

datumaro/datumaro/plugins/tf_detection_api_format/extractor.py

@@ -145,9 +145,7 @@ class TfDetectionApiExtractor(SourceExtractor):
                     continue
                 dataset_labels[label] = label_id - 1
 
-            item_id = frame_id
-            if not item_id:
-                item_id = osp.splitext(frame_filename)[0]
+            item_id = osp.splitext(frame_filename)[0]
 
             annotations = []
             for shape_id, shape in enumerate(
@@ -188,6 +186,7 @@ class TfDetectionApiExtractor(SourceExtractor):
                 image = Image(**image_params, size=image_size)
 
             dataset_items.append(DatasetItem(id=item_id, subset=subset,
-                image=image, annotations=annotations))
+                image=image, annotations=annotations,
+                attributes={'source_id': frame_id}))
 
         return dataset_items, dataset_labels

datumaro/datumaro/plugins/transforms.py

@@ -364,10 +364,13 @@ class RandomSplit(Transform, CliPlugin):
 
 class IdFromImageName(Transform, CliPlugin):
     def transform_item(self, item):
-        name = item.id
-        if item.has_image and item.image.filename:
-            name = osp.splitext(item.image.filename)[0]
-        return self.wrap_item(item, id=name)
+        if item.has_image and item.image.path:
+            name = osp.splitext(osp.basename(item.image.path))[0]
+            return self.wrap_item(item, id=name)
+        else:
+            log.debug("Can't change item id for item '%s': "
+                "item has no image info" % item.id)
+            return item
 
 class RemapLabels(Transform, CliPlugin):
     DefaultAction = Enum('DefaultAction', ['keep', 'delete'])

datumaro/datumaro/plugins/voc_format/converter.py

@@ -135,16 +135,9 @@ class _Converter:
             for item in subset:
                 log.debug("Converting item '%s'", item.id)
 
-                image_filename = ''
-                if item.has_image:
-                    image_filename = item.image.filename
+                image_filename = item.id + VocPath.IMAGE_EXT
                 if self._save_images:
                     if item.has_image and item.image.has_data:
-                        if image_filename:
-                            image_filename = osp.splitext(image_filename)[0]
-                        else:
-                            image_filename = item.id
-                        image_filename += VocPath.IMAGE_EXT
                         save_image(osp.join(self._images_dir, image_filename),
                             item.image.data, create_dir=True)
                     else:

datumaro/datumaro/plugins/voc_format/extractor.py

@@ -108,8 +108,8 @@ class _VocXmlExtractor(_VocExtractor):
 
         for item_id in self._items:
             log.debug("Reading item '%s'" % item_id)
-            image = osp.join(self._dataset_dir, VocPath.IMAGES_DIR,
-                item_id + VocPath.IMAGE_EXT)
+            image = item_id + VocPath.IMAGE_EXT
+            height, width = 0, 0
 
             anns = []
             ann_file = osp.join(anno_dir, item_id + '.xml')
@@ -121,11 +121,15 @@ class _VocXmlExtractor(_VocExtractor):
                 width = root_elem.find('size/width')
                 if width is not None:
                     width = int(width.text)
-                if height and width:
-                    image = Image(path=image, size=(height, width))
-
+                filename_elem = root_elem.find('filename')
+                if filename_elem is not None:
+                    image = filename_elem.text
                 anns = self._parse_annotations(root_elem)
 
+            image = osp.join(self._dataset_dir, VocPath.IMAGES_DIR, image)
+            if height and width:
+                image = Image(path=image, size=(height, width))
+
             yield DatasetItem(id=item_id, subset=self._subset,
                 image=image, annotations=anns)
 

datumaro/datumaro/plugins/yolo_format/converter.py

@@ -80,13 +80,9 @@ class YoloConverter(Converter, CliPlugin):
                         "item has no image info" % item.id)
                 height, width = item.image.size
 
-                image_name = item.image.filename
-                item_name = osp.splitext(item.image.filename)[0]
+                image_name = item.id + '.jpg'
                 if self._save_images:
                     if item.has_image and item.image.has_data:
-                        if not item_name:
-                            item_name = item.id
-                        image_name = item_name + '.jpg'
                         save_image(osp.join(subset_dir, image_name),
                             item.image.data, create_dir=True)
                     else:
@@ -105,7 +101,8 @@ class YoloConverter(Converter, CliPlugin):
                     yolo_bb = ' '.join('%.6f' % p for p in yolo_bb)
                     yolo_annotation += '%s %s\n' % (bbox.label, yolo_bb)
 
-                annotation_path = osp.join(subset_dir, '%s.txt' % item_name)
+                annotation_path = osp.join(subset_dir, '%s.txt' % item.id)
+                os.makedirs(osp.dirname(annotation_path), exist_ok=True)
                 with open(annotation_path, 'w') as f:
                     f.write(yolo_annotation)
 

datumaro/datumaro/plugins/yolo_format/extractor.py

@@ -10,6 +10,7 @@ import re
 from datumaro.components.extractor import (SourceExtractor, Extractor,
     DatasetItem, AnnotationType, Bbox, LabelCategories
 )
+from datumaro.util import split_path
 from datumaro.util.image import Image
 
 from .format import YoloPath
@@ -83,14 +84,14 @@ class YoloExtractor(SourceExtractor):
                 config_path)
 
         for subset_name, list_path in subsets.items():
-            list_path = self._make_local_path(list_path)
+            list_path = osp.join(self._path, self.localize_path(list_path))
             if not osp.isfile(list_path):
                 raise Exception("Not found '%s' subset list file" % subset_name)
 
             subset = YoloExtractor.Subset(subset_name, self)
             with open(list_path, 'r') as f:
                 subset.items = OrderedDict(
-                    (osp.splitext(osp.basename(p.strip()))[0], p.strip())
+                    (self.name_from_path(p), self.localize_path(p))
                     for p in f
                 )
             subsets[subset_name] = subset
@@ -99,25 +100,38 @@ class YoloExtractor(SourceExtractor):
 
         self._categories = {
             AnnotationType.label:
-                self._load_categories(self._make_local_path(names_path))
+                self._load_categories(
+                    osp.join(self._path, self.localize_path(names_path)))
         }
 
-    def _make_local_path(self, path):
+    @staticmethod
+    def localize_path(path):
+        path = path.strip()
         default_base = osp.join('data', '')
         if path.startswith(default_base): # default path
             path = path[len(default_base) : ]
-        return osp.join(self._path, path) # relative or absolute path
+        return path
+
+    @classmethod
+    def name_from_path(cls, path):
+        path = cls.localize_path(path)
+        parts = split_path(path)
+        if 1 < len(parts) and not osp.isabs(path):
+            # NOTE: when path is like [data/]<subset_obj>/<image_name>
+            # drop everything but <image name>
+            # <image name> can be <a/b/c/filename.ext>, so no just basename()
+            path = osp.join(*parts[1:])
+        return osp.splitext(path)[0]
 
     def _get(self, item_id, subset_name):
         subset = self._subsets[subset_name]
         item = subset.items[item_id]
 
         if isinstance(item, str):
-            image_path = self._make_local_path(item)
             image_size = self._image_info.get(item_id)
-            image = Image(path=image_path, size=image_size)
+            image = Image(path=osp.join(self._path, item), size=image_size)
 
-            anno_path = osp.splitext(image_path)[0] + '.txt'
+            anno_path = osp.splitext(image.path)[0] + '.txt'
             annotations = self._parse_annotations(anno_path, image)
 
             item = DatasetItem(id=item_id, subset=subset_name,
@@ -137,8 +151,10 @@ class YoloExtractor(SourceExtractor):
 
         annotations = []
         if lines:
-            # use image info as late as possible
-            image_height, image_width = image.size
+            size = image.size # use image info as late as possible
+            if size is None:
+                raise Exception("Can't find image info for '%s'" % image.path)
+            image_height, image_width = size
         for line in lines:
             label_id, xc, yc, w, h = line.split()
             label_id = int(label_id)

datumaro/datumaro/util/image.py

@@ -215,10 +215,6 @@ class Image:
     def path(self):
         return self._path
 
-    @property
-    def filename(self):
-        return osp.basename(self._path)
-
     @property
     def data(self):
         if callable(self._data):

datumaro/datumaro/util/test_utils.py

@@ -87,6 +87,7 @@ def compare_datasets(test, expected, actual):
         item_b = find(actual, lambda x: x.id == item_a.id and \
             x.subset == item_a.subset)
         test.assertFalse(item_b is None, item_a.id)
+        test.assertEqual(item_a.attributes, item_b.attributes)
         test.assertEqual(len(item_a.annotations), len(item_b.annotations))
         for ann_a in item_a.annotations:
             # We might find few corresponding items, so check them all

datumaro/tests/assets/coco_dataset/annotations/instances_val.json

@@ -56,4 +56,4 @@
         "iscrowd": 1
       }
     ]
-  }
+  }

datumaro/tests/assets/voc_dataset/Annotations/2007_000001.xml

@@ -51,4 +51,4 @@
       <walking>1</walking>
     </actions>
   </object>
-</annotation>
+</annotation>

datumaro/tests/test_coco_format.py

@@ -28,7 +28,8 @@ class CocoImporterTest(TestCase):
         class DstExtractor(Extractor):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, image=np.ones((10, 5, 3)), subset='val',
+                    DatasetItem(id='000000000001', image=np.ones((10, 5, 3)),
+                        subset='val', attributes={'id': 1},
                         annotations=[
                             Polygon([0, 0, 1, 0, 1, 2, 0, 2], label=0,
                                 id=1, group=1, attributes={'is_crowd': False}),
@@ -76,16 +77,16 @@ class CocoConverterTest(TestCase):
                         annotations=[
                             Caption('hello', id=1, group=1),
                             Caption('world', id=2, group=2),
-                        ]),
+                        ], attributes={'id': 1}),
                     DatasetItem(id=2, subset='train',
                         annotations=[
                             Caption('test', id=3, group=3),
-                        ]),
+                        ], attributes={'id': 2}),
 
                     DatasetItem(id=3, subset='val',
                         annotations=[
                             Caption('word', id=1, group=1),
-                        ]
+                        ], attributes={'id': 1}
                     ),
                 ])
 
@@ -111,7 +112,7 @@ class CocoConverterTest(TestCase):
                             Polygon([0, 1, 2, 1, 2, 3, 0, 3],
                                 attributes={ 'is_crowd': False },
                                 label=2, group=1, id=1),
-                        ]),
+                        ], attributes={'id': 1}),
                     DatasetItem(id=2, subset='train', image=np.ones((4, 4, 3)),
                         annotations=[
                             # Mask + bbox
@@ -125,7 +126,7 @@ class CocoConverterTest(TestCase):
                                 label=4, group=3, id=3),
                             Bbox(1, 0, 2, 2, label=4, group=3, id=3,
                                 attributes={ 'is_crowd': True }),
-                        ]),
+                        ], attributes={'id': 2}),
 
                     DatasetItem(id=3, subset='val', image=np.ones((4, 4, 3)),
                         annotations=[
@@ -140,7 +141,7 @@ class CocoConverterTest(TestCase):
                                 ),
                                 attributes={ 'is_crowd': True },
                                 label=4, group=3, id=3),
-                        ]),
+                        ], attributes={'id': 1}),
                 ])
 
             def categories(self):
@@ -154,7 +155,7 @@ class CocoConverterTest(TestCase):
                             Polygon([0, 1, 2, 1, 2, 3, 0, 3],
                                 attributes={ 'is_crowd': False },
                                 label=2, group=1, id=1),
-                        ]),
+                        ], attributes={'id': 1}),
                     DatasetItem(id=2, subset='train', image=np.ones((4, 4, 3)),
                         annotations=[
                             Mask(np.array([
@@ -165,7 +166,7 @@ class CocoConverterTest(TestCase):
                                 ),
                                 attributes={ 'is_crowd': True },
                                 label=4, group=3, id=3),
-                        ]),
+                        ], attributes={'id': 2}),
 
                     DatasetItem(id=3, subset='val', image=np.ones((4, 4, 3)),
                         annotations=[
@@ -177,7 +178,7 @@ class CocoConverterTest(TestCase):
                                 ),
                                 attributes={ 'is_crowd': True },
                                 label=4, group=3, id=3),
-                        ]),
+                        ], attributes={'id': 1}),
                 ])
 
             def categories(self):
@@ -227,7 +228,7 @@ class CocoConverterTest(TestCase):
                             ),
                             label=3, id=4, group=4,
                             attributes={ 'is_crowd': False }),
-                        ]
+                        ], attributes={'id': 1}
                     ),
                 ])
 
@@ -285,7 +286,7 @@ class CocoConverterTest(TestCase):
                             Polygon([1, 1, 4, 1, 4, 4, 1, 4],
                                 label=1, id=2, group=2,
                                 attributes={ 'is_crowd': False }),
-                        ]
+                        ], attributes={'id': 1}
                     ),
                 ])
 
@@ -335,7 +336,7 @@ class CocoConverterTest(TestCase):
                                 ),
                                 attributes={ 'is_crowd': True },
                                 label=3, id=4, group=4),
-                        ]
+                        ], attributes={'id': 1}
                     ),
                 ])
 
@@ -385,7 +386,7 @@ class CocoConverterTest(TestCase):
                                 [5.0, 3.5, 4.5, 0.0, 8.0, 0.0, 5.0, 3.5],
                                 label=3, id=4, group=4,
                                 attributes={ 'is_crowd': False }),
-                        ]
+                        ], attributes={'id': 1}
                     ),
                 ])
 
@@ -401,14 +402,14 @@ class CocoConverterTest(TestCase):
         class TestExtractor(Extractor):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='train'),
-                    DatasetItem(id=2, subset='train'),
+                    DatasetItem(id=1, subset='train', attributes={'id': 1}),
+                    DatasetItem(id=2, subset='train', attributes={'id': 2}),
 
-                    DatasetItem(id=2, subset='val'),
-                    DatasetItem(id=3, subset='val'),
-                    DatasetItem(id=4, subset='val'),
+                    DatasetItem(id=2, subset='val', attributes={'id': 2}),
+                    DatasetItem(id=3, subset='val', attributes={'id': 3}),
+                    DatasetItem(id=4, subset='val', attributes={'id': 4}),
 
-                    DatasetItem(id=5, subset='test'),
+                    DatasetItem(id=5, subset='test', attributes={'id': 1}),
                 ])
 
         with TestDir() as test_dir:
@@ -423,7 +424,7 @@ class CocoConverterTest(TestCase):
                         annotations=[
                             Label(4, id=1, group=1),
                             Label(9, id=2, group=2),
-                        ]
+                        ], attributes={'id': 1}
                     ),
                 ])
 
@@ -511,7 +512,7 @@ class CocoConverterTest(TestCase):
                             Polygon([1, 2, 3, 2, 3, 4, 1, 4],
                                 group=5, id=5,
                                 attributes={'is_crowd': False}),
-                        ]),
+                        ], attributes={'id': 1}),
                 ])
 
         with TestDir() as test_dir:
@@ -523,8 +524,8 @@ class CocoConverterTest(TestCase):
         class TestExtractor(Extractor):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1),
-                    DatasetItem(id=2),
+                    DatasetItem(id=1, attributes={'id': 1}),
+                    DatasetItem(id=2, attributes={'id': 2}),
                 ])
 
             def categories(self):
@@ -538,9 +539,38 @@ class CocoConverterTest(TestCase):
         class TestExtractor(Extractor):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, image=Image(path='1.jpg', size=(10, 15))),
+                    DatasetItem(id=1, image=Image(path='1.jpg', size=(10, 15)),
+                        attributes={'id': 1}),
+                ])
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(TestExtractor(),
+                CocoConverter(tasks='image_info'), test_dir)
+
+    def test_relative_paths(self):
+        class TestExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='1', image=np.ones((4, 2, 3)),
+                        attributes={'id': 1}),
+                    DatasetItem(id='subdir1/1', image=np.ones((2, 6, 3)),
+                        attributes={'id': 2}),
+                    DatasetItem(id='subdir2/1', image=np.ones((5, 4, 3)),
+                        attributes={'id': 3}),
+                ])
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(TestExtractor(),
+                CocoConverter(tasks='image_info', save_images=True), test_dir)
+
+    def test_preserve_coco_ids(self):
+        class TestExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='some/name1', image=np.ones((4, 2, 3)),
+                        attributes={'id': 40}),
                 ])
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoConverter(tasks='image_info'), test_dir)
+                CocoConverter(tasks='image_info', save_images=True), test_dir)

datumaro/tests/test_cvat_format.py

@@ -30,7 +30,7 @@ class CvatImporterTest(TestCase):
         class DstExtractor(Extractor):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=0, subset='train',
+                    DatasetItem(id='img0', subset='train',
                         image=np.ones((8, 8, 3)),
                         annotations=[
                             Bbox(0, 2, 4, 2, label=0, z_order=1,
@@ -40,15 +40,15 @@ class CvatImporterTest(TestCase):
                                 }),
                             PolyLine([1, 2, 3, 4, 5, 6, 7, 8],
                                 attributes={'occluded': False}),
-                        ]),
-                    DatasetItem(id=1, subset='train',
+                        ], attributes={'frame': 0}),
+                    DatasetItem(id='img1', subset='train',
                         image=np.ones((10, 10, 3)),
                         annotations=[
                             Polygon([1, 2, 3, 4, 6, 5], z_order=1,
                                 attributes={'occluded': False}),
                             Points([1, 2, 3, 4, 5, 6], label=1, z_order=2,
                                 attributes={'occluded': False}),
-                        ]),
+                        ], attributes={'frame': 1}),
                 ])
 
             def categories(self):
@@ -65,7 +65,7 @@ class CvatImporterTest(TestCase):
         class DstExtractor(Extractor):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=10, subset='annotations',
+                    DatasetItem(id='frame_000010', subset='annotations',
                         image=np.ones((20, 25, 3)),
                         annotations=[
                             Bbox(3, 4, 7, 1, label=2,
@@ -83,8 +83,8 @@ class CvatImporterTest(TestCase):
                                     'outside': False, 'keyframe': True,
                                     'track_id': 1, 'hgl': 'hgkf',
                                 }),
-                        ]),
-                    DatasetItem(id=13, subset='annotations',
+                        ], attributes={'frame': 10}),
+                    DatasetItem(id='frame_000013', subset='annotations',
                         image=np.ones((20, 25, 3)),
                         annotations=[
                             Bbox(7, 6, 7, 2, label=2,
@@ -110,10 +110,9 @@ class CvatImporterTest(TestCase):
                                     'outside': False, 'keyframe': True,
                                     'track_id': 2,
                                 }),
-                        ]),
-                    DatasetItem(id=16, subset='annotations',
-                        image=Image(path='frame_0000016.png',
-                            size=(20, 25)), # no image in the dataset files
+                        ], attributes={'frame': 13}),
+                    DatasetItem(id='frame_000016', subset='annotations',
+                        image=Image(path='frame_0000016.png', size=(20, 25)),
                         annotations=[
                             Bbox(8, 7, 6, 10, label=2,
                                 id=0,
@@ -130,7 +129,7 @@ class CvatImporterTest(TestCase):
                                     'outside': True, 'keyframe': True,
                                     'track_id': 2,
                                 }),
-                        ]),
+                        ], attributes={'frame': 16}),
                 ])
 
             def categories(self):
@@ -220,7 +219,7 @@ class CvatConverterTest(TestCase):
                                     'a1': 'x', 'a2': 42 }),
                             Label(1),
                             Label(2, attributes={ 'a1': 'y', 'a2': 44 }),
-                        ]
+                        ], attributes={'frame': 0}
                     ),
                     DatasetItem(id=1, subset='s1',
                         annotations=[
@@ -230,7 +229,7 @@ class CvatConverterTest(TestCase):
                             Bbox(5, 0, 1, 9,
                                 label=3, group=4,
                                 attributes={ 'occluded': False }),
-                        ]
+                        ], attributes={'frame': 1}
                     ),
 
                     DatasetItem(id=2, subset='s2', image=np.ones((5, 10, 3)),
@@ -238,11 +237,12 @@ class CvatConverterTest(TestCase):
                             Polygon([0, 0, 4, 0, 4, 4], z_order=1,
                                 label=3, group=4,
                                 attributes={ 'occluded': False }),
-                        ]
+                        ], attributes={'frame': 0}
                     ),
 
                     DatasetItem(id=3, subset='s3', image=Image(
-                        path='3.jpg', size=(2, 4))),
+                            path='3.jpg', size=(2, 4)),
+                        attributes={'frame': 0}),
                 ])
 
             def categories(self):
@@ -252,3 +252,49 @@ class CvatConverterTest(TestCase):
             self._test_save_and_load(SrcExtractor(),
                 CvatConverter(save_images=True), test_dir,
                 target_dataset=DstExtractor())
+
+    def test_relative_paths(self):
+        class SrcExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='1', image=np.ones((4, 2, 3))),
+                    DatasetItem(id='subdir1/1', image=np.ones((2, 6, 3))),
+                    DatasetItem(id='subdir2/1', image=np.ones((5, 4, 3))),
+                ])
+
+            def categories(self):
+                return { AnnotationType.label: LabelCategories() }
+
+        class DstExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='1', image=np.ones((4, 2, 3)),
+                        attributes={'frame': 0}),
+                    DatasetItem(id='subdir1/1', image=np.ones((2, 6, 3)),
+                        attributes={'frame': 1}),
+                    DatasetItem(id='subdir2/1', image=np.ones((5, 4, 3)),
+                        attributes={'frame': 2}),
+                ])
+
+            def categories(self):
+                return { AnnotationType.label: LabelCategories() }
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(SrcExtractor(),
+                CvatConverter(save_images=True), test_dir,
+                target_dataset=DstExtractor())
+
+    def test_preserve_frame_ids(self):
+        class TestExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='some/name1', image=np.ones((4, 2, 3)),
+                        attributes={'frame': 40}),
+                ])
+
+            def categories(self):
+                return { AnnotationType.label: LabelCategories() }
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(TestExtractor(),
+                CvatConverter(save_images=True), test_dir)

datumaro/tests/test_datumaro_format.py

@@ -14,6 +14,7 @@ from datumaro.util.mask_tools import generate_colormap
 from datumaro.util.image import Image
 from datumaro.util.test_utils import TestDir, compare_datasets_strict
 
+
 class DatumaroConverterTest(TestCase):
     def _test_save_and_load(self, source_dataset, converter, test_dir,
             target_dataset=None, importer_args=None):
@@ -96,3 +97,16 @@ class DatumaroConverterTest(TestCase):
             DatumaroConverter()(self.TestExtractor(), save_dir=test_dir)
 
             self.assertTrue(DatumaroImporter.detect(test_dir))
+
+    def test_relative_paths(self):
+        class TestExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='1', image=np.ones((4, 2, 3))),
+                    DatasetItem(id='subdir1/1', image=np.ones((2, 6, 3))),
+                    DatasetItem(id='subdir2/1', image=np.ones((5, 4, 3))),
+                ])
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(TestExtractor(),
+                DatumaroConverter(save_images=True), test_dir)

datumaro/tests/test_image_dir_format.py

@@ -26,3 +26,23 @@ class ImageDirFormatTest(TestCase):
             parsed_dataset = project.make_dataset()
 
             compare_datasets(self, source_dataset, parsed_dataset)
+
+    def test_relative_paths(self):
+        class TestExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='1', image=np.ones((4, 2, 3))),
+                    DatasetItem(id='subdir1/1', image=np.ones((2, 6, 3))),
+                    DatasetItem(id='subdir2/1', image=np.ones((5, 4, 3))),
+                ])
+
+        with TestDir() as test_dir:
+            source_dataset = TestExtractor()
+
+            ImageDirConverter()(source_dataset, save_dir=test_dir)
+
+            project = Project.import_from(test_dir, 'image_dir')
+            parsed_dataset = project.make_dataset()
+
+            compare_datasets(self, source_dataset, parsed_dataset)
+

datumaro/tests/test_labelme_format.py

@@ -108,6 +108,21 @@ class LabelMeConverterTest(TestCase):
                 SrcExtractor(), LabelMeConverter(save_images=True),
                 test_dir, target_dataset=DstExtractor())
 
+    def test_cant_save_dataset_with_relative_paths(self):
+        class SrcExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='dir/1', image=np.ones((2, 6, 3))),
+                ])
+
+            def categories(self):
+                return { AnnotationType.label: LabelCategories() }
+
+        with self.assertRaisesRegex(Exception, r'only supports flat'):
+            with TestDir() as test_dir:
+                self._test_save_and_load(SrcExtractor(),
+                    LabelMeConverter(save_images=True), test_dir)
+
 
 DUMMY_DATASET_DIR = osp.join(osp.dirname(__file__), 'assets', 'labelme_dataset')
 

datumaro/tests/test_tfrecord_format.py

@@ -56,7 +56,7 @@ class TfrecordConverterTest(TestCase):
                             Bbox(0, 4, 4, 8, label=2),
                             Bbox(0, 4, 4, 4, label=3),
                             Bbox(2, 4, 4, 4),
-                        ]
+                        ], attributes={'source_id': ''}
                     ),
                 ])
 
@@ -85,7 +85,8 @@ class TfrecordConverterTest(TestCase):
                                 [0, 1, 1, 0],
                                 [1, 0, 0, 1],
                             ]), label=1),
-                        ]
+                        ],
+                        attributes={'source_id': ''}
                     ),
                 ])
 
@@ -111,18 +112,21 @@ class TfrecordConverterTest(TestCase):
                         annotations=[
                             Bbox(2, 1, 4, 4, label=2),
                             Bbox(4, 2, 8, 4, label=3),
-                        ]
+                        ],
+                        attributes={'source_id': ''}
                     ),
 
                     DatasetItem(id=2,
                         image=np.ones((8, 8, 3)) * 2,
                         annotations=[
                             Bbox(4, 4, 4, 4, label=3),
-                        ]
+                        ],
+                        attributes={'source_id': ''}
                     ),
 
                     DatasetItem(id=3,
                         image=np.ones((8, 4, 3)) * 3,
+                        attributes={'source_id': ''}
                     ),
                 ])
 
@@ -143,7 +147,10 @@ class TfrecordConverterTest(TestCase):
         class TestExtractor(Extractor):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, image=Image(path='1/q.e', size=(10, 15))),
+                    DatasetItem(id='1/q.e',
+                        image=Image(path='1/q.e', size=(10, 15)),
+                        attributes={'source_id': ''}
+                    )
                 ])
 
             def categories(self):
@@ -199,6 +206,7 @@ class TfrecordImporterTest(TestCase):
                             Bbox(0, 4, 4, 4, label=3),
                             Bbox(2, 4, 4, 4),
                         ],
+                        attributes={'source_id': '1'}
                     ),
 
                     DatasetItem(id=2, subset='val',
@@ -206,10 +214,12 @@ class TfrecordImporterTest(TestCase):
                         annotations=[
                             Bbox(1, 2, 4, 2, label=3),
                         ],
+                        attributes={'source_id': '2'}
                     ),
 
                     DatasetItem(id=3, subset='test',
                         image=np.ones((5, 4, 3)) * 3,
+                        attributes={'source_id': '3'}
                     ),
                 ])
 

datumaro/tests/test_voc_format.py

@@ -145,7 +145,7 @@ class VocConverterTest(TestCase):
         class TestExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=0, subset='a', annotations=[
+                    DatasetItem(id='a/0', subset='a', annotations=[
                         Label(1),
                         Label(2),
                         Label(3),
@@ -164,7 +164,7 @@ class VocConverterTest(TestCase):
         class TestExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='a', annotations=[
+                    DatasetItem(id='a/1', subset='a', annotations=[
                         Bbox(2, 3, 4, 5, label=2,
                             attributes={ 'occluded': True }
                         ),
@@ -183,7 +183,7 @@ class VocConverterTest(TestCase):
         class DstExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='a', annotations=[
+                    DatasetItem(id='a/1', subset='a', annotations=[
                         Bbox(2, 3, 4, 5, label=2, id=1, group=1,
                             attributes={
                                 'truncated': False,
@@ -220,7 +220,7 @@ class VocConverterTest(TestCase):
         class TestExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='a', annotations=[
+                    DatasetItem(id='a/b/1', subset='a', annotations=[
                         # overlapping masks, the first should be truncated
                         # the second and third are different instances
                         Mask(image=np.array([[0, 0, 0, 1, 0]]), label=3,
@@ -235,7 +235,7 @@ class VocConverterTest(TestCase):
         class DstExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='a', annotations=[
+                    DatasetItem(id='a/b/1', subset='a', annotations=[
                         Mask(image=np.array([[0, 0, 1, 0, 0]]), label=4,
                             group=1),
                         Mask(image=np.array([[1, 1, 0, 0, 0]]), label=3,
@@ -323,7 +323,7 @@ class VocConverterTest(TestCase):
         class TestExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='a', annotations=[
+                    DatasetItem(id='a/b/1', subset='a', annotations=[
                         Bbox(2, 3, 4, 5, label=2, id=1, group=1,
                             attributes={
                                 'pose': VOC.VocPose(1).name,
@@ -347,7 +347,7 @@ class VocConverterTest(TestCase):
         class TestExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='a', annotations=[
+                    DatasetItem(id='a/b/1', subset='a', annotations=[
                         Bbox(2, 3, 4, 5, label=2,
                             attributes={
                                 'truncated': True,
@@ -368,7 +368,7 @@ class VocConverterTest(TestCase):
         class DstExtractor(TestExtractorBase):
             def __iter__(self):
                 return iter([
-                    DatasetItem(id=1, subset='a', annotations=[
+                    DatasetItem(id='a/b/1', subset='a', annotations=[
                         Bbox(2, 3, 4, 5, label=2,
                             id=1, group=1, attributes={
                                 'truncated': True,
@@ -666,3 +666,16 @@ class VocConverterTest(TestCase):
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
                 VocConverter(label_map='voc'), test_dir)
+
+    def test_relative_paths(self):
+        class TestExtractor(TestExtractorBase):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='1', image=np.ones((4, 2, 3))),
+                    DatasetItem(id='subdir1/1', image=np.ones((2, 6, 3))),
+                    DatasetItem(id='subdir2/1', image=np.ones((5, 4, 3))),
+                ])
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(TestExtractor(),
+                VocConverter(label_map='voc', save_images=True), test_dir)

datumaro/tests/test_yolo_format.py

@@ -116,6 +116,32 @@ class YoloFormatTest(TestCase):
 
             compare_datasets(self, source_dataset, parsed_dataset)
 
+    def test_relative_paths(self):
+        class TestExtractor(Extractor):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id='1', subset='train',
+                        image=np.ones((4, 2, 3))),
+                    DatasetItem(id='subdir1/1', subset='train',
+                        image=np.ones((2, 6, 3))),
+                    DatasetItem(id='subdir2/1', subset='train',
+                        image=np.ones((5, 4, 3))),
+                ])
+
+            def categories(self):
+                return { AnnotationType.label: LabelCategories() }
+
+        for save_images in {True, False}:
+            with self.subTest(save_images=save_images):
+                with TestDir() as test_dir:
+                    source_dataset = TestExtractor()
+
+                    YoloConverter(save_images=save_images)(
+                        source_dataset, test_dir)
+                    parsed_dataset = YoloImporter()(test_dir).make_dataset()
+
+                    compare_datasets(self, source_dataset, parsed_dataset)
+
 
 DUMMY_DATASET_DIR = osp.join(osp.dirname(__file__), 'assets', 'yolo_dataset')