Fixed VGGFace2

CHANGELOG.md

@@ -11,8 +11,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `WiderFace` dataset format (<https://github.com/openvinotoolkit/datumaro/pull/65>)
 - Function to transform annotations to labels (<https://github.com/openvinotoolkit/datumaro/pull/66>)
 - Task-specific Splitter (<https://github.com/openvinotoolkit/datumaro/pull/68>, <https://github.com/openvinotoolkit/datumaro/pull/81>)
-- `VGGFace2` dataset format (<https://github.com/openvinotoolkit/datumaro/pull/69>)
-- Unique image count statistic (<https://github.com/openvinotoolkit/datumaro/pull/87)
+- `VGGFace2` dataset format (<https://github.com/openvinotoolkit/datumaro/pull/69>, <https://github.com/openvinotoolkit/datumaro/pull/82>)
+- Unique image count statistic (<https://github.com/openvinotoolkit/datumaro/pull/87>)
 
 ### Changed
 - `Dataset` class extended with new operations: `save`, `load`, `export`, `import_from`, `detect`, `run_model` (<https://github.com/openvinotoolkit/datumaro/pull/71>)

datumaro/plugins/vgg_face2_format.py

@@ -8,14 +8,15 @@
 
 from datumaro.components.converter import Converter
 from datumaro.components.extractor import (AnnotationType, Bbox, DatasetItem,
-    Importer, Points, LabelCategories, SourceExtractor)
+    Importer, Label, LabelCategories, Points, SourceExtractor)
 
 
 class VggFace2Path:
     ANNOTATION_DIR = "bb_landmark"
     IMAGE_EXT = '.jpg'
     BBOXES_FILE = 'loose_bb_'
     LANDMARKS_FILE = 'loose_landmark_'
+    LABELS_FILE = 'labels.txt'
 
 class VggFace2Extractor(SourceExtractor):
     def __init__(self, path):
@@ -29,11 +30,29 @@ def __init__(self, path):
             subset = subset.split('_')[2]
         super().__init__(subset=subset)
 
-        self._load_categories()
+        self._categories = self._load_categories()
         self._items = list(self._load_items(path).values())
 
     def _load_categories(self):
-        self._categories[AnnotationType.label] = LabelCategories()
+        label_cat = LabelCategories()
+        path = osp.join(self._dataset_dir, VggFace2Path.LABELS_FILE)
+        if osp.isfile(path):
+            with open(path, encoding='utf-8') as labels_file:
+                lines = [s.strip() for s in labels_file]
+            for line in lines:
+                objects = line.split()
+                label = objects[0]
+                class_name = None
+                if 1 < len(objects):
+                    class_name = objects[1]
+                label_cat.add(label, parent=class_name)
+        else:
+            subset_path = osp.join(self._dataset_dir, self._subset)
+            if osp.isdir(subset_path):
+                for images_dir in sorted(os.listdir(subset_path)):
+                    if osp.isdir(osp.join(subset_path, images_dir)):
+                       label_cat.add(images_dir)
+        return { AnnotationType.label: label_cat }
 
     def _load_items(self, path):
         items = {}
@@ -42,30 +61,47 @@ def _load_items(self, path):
 
         for row in landmarks_table:
             item_id = row['NAME_ID']
-            image_path = osp.join(self._dataset_dir, self._subset,
-                item_id + VggFace2Path.IMAGE_EXT)
-            annotations = []
+            label = None
+            if '/' in item_id:
+                label_name = item_id.split('/')[0]
+                label = self._categories[AnnotationType.label].find(label_name)[0]
+                if label is not None:
+                    item_id = item_id[len(label_name) + 1:]
+            if item_id not in items:
+                image_path = osp.join(self._dataset_dir, self._subset,
+                    row['NAME_ID'] + VggFace2Path.IMAGE_EXT)
+                items[item_id] = DatasetItem(id=item_id, subset=self._subset,
+                    image=image_path)
+            annotations = items[item_id].annotations
             if len([p for p in row if row[p] == '']) == 0 and len(row) == 11:
                 annotations.append(Points(
-                    [float(row[p]) for p in row if p != 'NAME_ID']))
-            if item_id in items and 0 < len(annotations):
-                annotation = items[item_id].annotations
-                annotation.append(annotations[0])
-            else:
-                items[item_id] = DatasetItem(id=item_id, subset=self._subset,
-                    image=image_path, annotations=annotations)
+                    [float(row[p]) for p in row if p != 'NAME_ID'], label=label,
+                    group=1))
+            elif label is not None:
+                annotations.append(Label(label=label, group=1))
 
         bboxes_path = osp.join(self._dataset_dir, VggFace2Path.ANNOTATION_DIR,
             VggFace2Path.BBOXES_FILE + self._subset + '.csv')
         if osp.isfile(bboxes_path):
             with open(bboxes_path) as content:
                 bboxes_table = list(csv.DictReader(content))
             for row in bboxes_table:
+                item_id = row['NAME_ID']
+                label = None
+                if '/' in item_id:
+                    label_name = item_id.split('/')[0]
+                    label = self._categories[AnnotationType.label].find(label_name)[0]
+                    if label is not None:
+                        item_id = item_id[len(label_name) + 1:]
+                if item_id not in items:
+                    image_path = osp.join(self._dataset_dir, self._subset,
+                        row['NAME_ID'] + VggFace2Path.IMAGE_EXT)
+                    items[item_id] = DatasetItem(id=item_id, subset=self._subset,
+                        image=image_path)
+                annotations = items[item_id].annotations
                 if len([p for p in row if row[p] == '']) == 0 and len(row) == 5:
-                    item_id = row['NAME_ID']
-                    annotations = items[item_id].annotations
-                    annotations.append(Bbox(int(row['X']), int(row['Y']),
-                        int(row['W']), int(row['H'])))
+                    annotations.append(Bbox(float(row['X']), float(row['Y']),
+                        float(row['W']), float(row['H']), label=label, group=1))
         return items
 
 class VggFace2Importer(Importer):
@@ -81,37 +117,76 @@ class VggFace2Converter(Converter):
 
     def apply(self):
         save_dir = self._save_dir
-
         os.makedirs(save_dir, exist_ok=True)
+
+        labels_path = osp.join(save_dir, VggFace2Path.LABELS_FILE)
+        labels_file = ''
+        for label in self._extractor.categories()[AnnotationType.label]:
+            labels_file += '%s' % label.name
+            if label.parent:
+                labels_file += ' %s' % label.parent
+            labels_file += '\n'
+        with open(labels_path, 'w', encoding='utf-8') as f:
+            f.write(labels_file)
+
+        label_categories = self._extractor.categories()[AnnotationType.label]
+
         for subset_name, subset in self._extractor.subsets().items():
             subset_dir = osp.join(save_dir, subset_name)
             bboxes_table = []
             landmarks_table = []
             for item in subset:
                 if item.has_image and self._save_images:
-                    self._save_image(item, osp.join(save_dir, subset_dir,
-                        item.id + VggFace2Path.IMAGE_EXT))
+                    labels = set(p.label for p in item.annotations
+                        if getattr(p, 'label') != None)
+                    if labels:
+                        for label in labels:
+                            self._save_image(item, osp.join(subset_dir,
+                                label_categories[label].name + '/' \
+                                + item.id + VggFace2Path.IMAGE_EXT))
+                    else:
+                        self._save_image(item, osp.join(subset_dir,
+                            item.id + VggFace2Path.IMAGE_EXT))
 
                 landmarks = [a for a in item.annotations
                     if a.type == AnnotationType.points]
-                if landmarks:
-                    for landmark in landmarks:
-                        points = landmark.points
-                        landmarks_table.append({'NAME_ID': item.id,
-                            'P1X': points[0], 'P1Y': points[1],
-                            'P2X': points[2], 'P2Y': points[3],
-                            'P3X': points[4], 'P3Y': points[5],
-                            'P4X': points[6], 'P4Y': points[7],
-                            'P5X': points[8], 'P5Y': points[9]})
-                else:
-                    landmarks_table.append({'NAME_ID': item.id})
+                for landmark in landmarks:
+                    name_id = item.id
+                    if landmark.label is not None \
+                        and label_categories[landmark.label].name:
+                        name_id = label_categories[landmark.label].name \
+                            + '/' + item.id
+                    points = landmark.points
+                    landmarks_table.append({'NAME_ID': name_id,
+                        'P1X': points[0], 'P1Y': points[1],
+                        'P2X': points[2], 'P2Y': points[3],
+                        'P3X': points[4], 'P3Y': points[5],
+                        'P4X': points[6], 'P4Y': points[7],
+                        'P5X': points[8], 'P5Y': points[9]})
 
                 bboxes = [a for a in item.annotations
                     if a.type == AnnotationType.bbox]
-                if bboxes:
-                    for bbox in bboxes:
-                        bboxes_table.append({'NAME_ID': item.id, 'X': int(bbox.x),
-                            'Y': int(bbox.y), 'W': int(bbox.w), 'H': int(bbox.h)})
+                for bbox in bboxes:
+                    name_id = item.id
+                    if bbox.label is not None and \
+                            label_categories[bbox.label].name:
+                        name_id = label_categories[bbox.label].name \
+                            + '/' + item.id
+                    bboxes_table.append({'NAME_ID': name_id, 'X': bbox.x,
+                        'Y': bbox.y, 'W': bbox.w, 'H': bbox.h})
+
+                labels = [a for a in item.annotations
+                    if a.type == AnnotationType.label]
+                for label in labels:
+                    name_id = item.id
+                    if label.label is not None and \
+                            label_categories[label.label].name:
+                        name_id = label_categories[label.label].name \
+                            + '/' + item.id
+                    landmarks_table.append({'NAME_ID': name_id})
+
+                if not landmarks and not bboxes and not labels:
+                    landmarks_table.append({'NAME_ID': item.id})
 
             landmarks_path = osp.join(save_dir, VggFace2Path.ANNOTATION_DIR,
                 VggFace2Path.LANDMARKS_FILE + subset_name + '.csv')

tests/assets/vgg_face2_dataset/labels.txt

@@ -0,0 +1,2 @@
+n000001 car
+n000002 person

tests/test_vgg_face2_format.py

@@ -2,8 +2,9 @@
 from unittest import TestCase
 
 import numpy as np
-from datumaro.components.extractor import Bbox, DatasetItem, Points
 from datumaro.components.dataset import Dataset
+from datumaro.components.extractor import (AnnotationType, Bbox, DatasetItem,
+    Label, LabelCategories, Points)
 from datumaro.plugins.vgg_face2_format import (VggFace2Converter,
     VggFace2Importer)
 from datumaro.util.test_utils import TestDir, compare_datasets
@@ -14,58 +15,78 @@ def test_can_save_and_load(self):
         source_dataset = Dataset.from_iterable([
             DatasetItem(id='1', subset='train', image=np.ones((8, 8, 3)),
                 annotations=[
-                    Bbox(0, 2, 4, 2),
+                    Bbox(0, 2, 4, 2, label=0, group=1),
                     Points([3.2, 3.12, 4.11, 3.2, 2.11,
-                        2.5, 3.5, 2.11, 3.8, 2.13]),
+                        2.5, 3.5, 2.11, 3.8, 2.13], label=0, group=1),
                 ]
             ),
             DatasetItem(id='2', subset='train', image=np.ones((10, 10, 3)),
                 annotations=[
                     Points([4.23, 4.32, 5.34, 4.45, 3.54,
-                        3.56, 4.52, 3.51, 4.78, 3.34]),
+                        3.56, 4.52, 3.51, 4.78, 3.34], label=1, group=1),
                 ]
             ),
-            DatasetItem(id='3', subset='val', image=np.ones((8, 8, 3))),
-            DatasetItem(id='4', subset='val', image=np.ones((10, 10, 3)),
+            DatasetItem(id='3', subset='train', image=np.ones((8, 8, 3)),
+                annotations=[Label(2, group=1)]
+            ),
+            DatasetItem(id='4', subset='train', image=np.ones((10, 10, 3)),
                 annotations=[
-                    Bbox(0, 2, 4, 2),
+                    Bbox(0, 2, 4, 2, label=3, group=1),
                     Points([3.2, 3.12, 4.11, 3.2, 2.11,
-                        2.5, 3.5, 2.11, 3.8, 2.13]),
-                    Bbox(2, 2, 1, 2),
-                    Points([2.787, 2.898, 2.965, 2.79, 2.8,
-                        2.456, 2.81, 2.32, 2.89, 2.3]),
+                        2.5, 3.5, 2.11, 3.8, 2.13], label=3, group=1),
                 ]
             ),
-            DatasetItem(id='5', subset='val', image=np.ones((8, 8, 3)),
+            DatasetItem(id='a/5', subset='train', image=np.ones((8, 8, 3)),
                 annotations=[
-                    Bbox(2, 2, 2, 2),
+                    Bbox(2, 2, 2, 2, group=1),
                 ]
             ),
-        ], categories=[])
+            DatasetItem(id='label_0', subset='train', image=np.ones((8, 8, 3)),
+            ),
+        ], categories={
+            AnnotationType.label: LabelCategories.from_iterable(
+                [('label_%s' % i, 'class_%s' % i) for i in range(5)]),
+        })
 
         with TestDir() as test_dir:
             VggFace2Converter.convert(source_dataset, test_dir, save_images=True)
-            parsed_dataset = VggFace2Importer()(test_dir).make_dataset()
+            parsed_dataset = Dataset.import_from(test_dir, 'vgg_face2')
 
             compare_datasets(self, source_dataset, parsed_dataset)
 
     def test_can_save_dataset_with_no_subsets(self):
         source_dataset = Dataset.from_iterable([
-            DatasetItem(id='a/b/1', image=np.ones((8, 8, 3)),
+            DatasetItem(id='b/1', image=np.ones((8, 8, 3)),
                 annotations=[
-                    Bbox(0, 2, 4, 2),
+                    Bbox(0, 2, 4, 2, label=0, group=1),
                     Points([4.23, 4.32, 5.34, 4.45, 3.54,
-                        3.56, 4.52, 3.51, 4.78, 3.34]),
+                        3.56, 4.52, 3.51, 4.78, 3.34], label=0, group=1),
                 ]
             ),
-        ], categories=[])
+        ], categories=['a'])
 
         with TestDir() as test_dir:
             VggFace2Converter.convert(source_dataset, test_dir, save_images=True)
-            parsed_dataset = VggFace2Importer()(test_dir).make_dataset()
+            parsed_dataset = Dataset.import_from(test_dir, 'vgg_face2')
 
             compare_datasets(self, source_dataset, parsed_dataset)
 
+    def test_can_save_dataset_with_no_save_images(self):
+        source_dataset = Dataset.from_iterable([
+            DatasetItem(id='1', image=np.ones((8, 8, 3)),
+                annotations=[
+                    Bbox(0, 2, 4, 2, label=0, group=1),
+                    Points([4.23, 4.32, 5.34, 4.45, 3.54,
+                        3.56, 4.52, 3.51, 4.78, 3.34], label=0, group=1),
+                ]
+            ),
+        ], categories=['label_0'])
+
+        with TestDir() as test_dir:
+            VggFace2Converter.convert(source_dataset, test_dir, save_images=False)
+            parsed_dataset = Dataset.import_from(test_dir, 'vgg_face2')
+
+            compare_datasets(self, source_dataset, parsed_dataset)
 
 DUMMY_DATASET_DIR = osp.join(osp.dirname(__file__), 'assets', 'vgg_face2_dataset')
 
@@ -75,23 +96,26 @@ def test_can_detect(self):
 
     def test_can_import(self):
         expected_dataset = Dataset.from_iterable([
-            DatasetItem(id='n000001/0001_01', subset='train',
+            DatasetItem(id='0001_01', subset='train',
                 image=np.ones((10, 15, 3)),
                 annotations=[
-                    Bbox(2, 2, 1, 2),
+                    Bbox(2, 2, 1, 2, label=0, group=1),
                     Points([2.787, 2.898, 2.965, 2.79, 2.8,
-                        2.456, 2.81, 2.32, 2.89, 2.3]),
+                        2.456, 2.81, 2.32, 2.89, 2.3], label=0, group=1),
                 ]
             ),
-            DatasetItem(id='n000002/0002_01', subset='train',
+            DatasetItem(id='0002_01', subset='train',
                 image=np.ones((10, 15, 3)),
                 annotations=[
-                    Bbox(1, 3, 1, 1),
+                    Bbox(1, 3, 1, 1, label=1, group=1),
                     Points([1.2, 3.8, 1.8, 3.82, 1.51,
-                        3.634, 1.43, 3.34, 1.65, 3.32])
+                        3.634, 1.43, 3.34, 1.65, 3.32], label=1, group=1)
                 ]
             ),
-        ], categories=[])
+        ], categories={
+            AnnotationType.label: LabelCategories.from_iterable(
+                [('n000001', 'car'), ('n000002', 'person')]),
+        })
 
         dataset = Dataset.import_from(DUMMY_DATASET_DIR, 'vgg_face2')