Add Wider Face format support

CHANGELOG.md

@@ -8,7 +8,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 ### Added
--
+- `WiderFace` dataset format (<https://github.com/openvinotoolkit/datumaro/pull/65>)
 
 ### Changed
 -

README.md

@@ -111,6 +111,7 @@ CVAT annotations                             ---> Publication, statistics etc.
   - [PASCAL VOC](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/htmldoc/index.html) (`classification`, `detection`, `segmentation`, `action_classification`, `person_layout`)
   - [YOLO](https://github.com/AlexeyAB/darknet#how-to-train-pascal-voc-data) (`bboxes`)
   - [TF Detection API](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/using_your_own_dataset.md) (`bboxes`, `masks`)
+  - [WIDER Face](http://shuoyang1213.me/WIDERFACE/) (`bboxes`)
   - [MOT sequences](https://arxiv.org/pdf/1906.04567.pdf)
   - [MOTS PNG](https://www.vision.rwth-aachen.de/page/mots)
   - [ImageNet](http://image-net.org/)

datumaro/plugins/widerface_format.py

@@ -0,0 +1,116 @@
+
+# Copyright (C) 2020 Intel Corporation
+#
+# SPDX-License-Identifier: MIT
+
+import os
+import os.path as osp
+import re
+
+from datumaro.components.converter import Converter
+from datumaro.components.extractor import (AnnotationType, Bbox, DatasetItem,
+    Importer, SourceExtractor)
+
+
+class WiderFacePath:
+    IMAGE_EXT = '.jpg'
+    ANNOTATIONS_DIR = 'wider_face_split'
+    IMAGES_DIR = 'images'
+    SUBSET_DIR = 'WIDER_'
+    BBOX_ATTRIBUTES = ['blur', 'expression', 'illumination',
+        'occlusion', 'pose', 'invalid']
+
+class WiderFaceExtractor(SourceExtractor):
+    def __init__(self, path):
+        if not osp.isfile(path):
+            raise Exception("Can't read annotation file '%s'" % path)
+        self._path = path
+        self._dataset_dir = osp.dirname(osp.dirname(path))
+
+        subset = osp.splitext(osp.basename(path))[0]
+        match = re.fullmatch(r'wider_face_\S+_bbx_gt', subset)
+        if match:
+            subset = subset.split('_')[2]
+        super().__init__(subset=subset)
+
+        self._items = list(self._load_items(path).values())
+
+    def _load_items(self, path):
+        items = {}
+        with open(path, 'r') as f:
+            lines = f.readlines()
+
+        image_ids = [image_id for image_id, line in enumerate(lines)
+            if WiderFacePath.IMAGE_EXT in line]
+
+        for image_id in image_ids:
+            image = lines[image_id]
+            image_path = osp.join(self._dataset_dir, WiderFacePath.SUBSET_DIR
+                + self._subset, WiderFacePath.IMAGES_DIR, image[:-1])
+            item_id = image[:-(len(WiderFacePath.IMAGE_EXT) + 1)]
+
+            bbox_count = lines[image_id + 1]
+            bbox_lines = lines[image_id + 2 : image_id + int(bbox_count) + 2]
+            annotations = []
+            for bbox in bbox_lines:
+                bbox_list = bbox.split()
+                if len(bbox_list) >= 4:
+                    attributes = {}
+                    if len(bbox_list) == 10:
+                        i = 4
+                        for attr in WiderFacePath.BBOX_ATTRIBUTES:
+                            attributes[attr] = int(bbox_list[i])
+                            i += 1
+                    annotations.append(Bbox(
+                        int(bbox_list[0]), int(bbox_list[1]),
+                        int(bbox_list[2]), int(bbox_list[3]),
+                        attributes = attributes
+                    ))
+
+            items[item_id] = DatasetItem(id=item_id, subset=self._subset,
+                image=image_path, annotations=annotations)
+        return items
+
+class WiderFaceImporter(Importer):
+    @classmethod
+    def find_sources(cls, path):
+        return cls._find_sources_recursive(osp.join(path,
+            WiderFacePath.ANNOTATIONS_DIR), '.txt', 'wider_face')
+
+class WiderFaceConverter(Converter):
+    DEFAULT_IMAGE_EXT = '.jpg'
+
+    def apply(self):
+        save_dir = self._save_dir
+
+        os.makedirs(save_dir, exist_ok=True)
+
+        for subset_name, subset in self._extractor.subsets().items():
+            subset_dir = osp.join(save_dir, WiderFacePath.SUBSET_DIR + subset_name)
+
+            wider_annotation = ''
+            for item in subset:
+                wider_annotation += '%s\n' % (item.id + WiderFacePath.IMAGE_EXT)
+                if item.has_image and self._save_images:
+                    self._save_image(item, osp.join(save_dir, subset_dir,
+                        WiderFacePath.IMAGES_DIR, item.id + WiderFacePath.IMAGE_EXT))
+
+                bboxes = [a for a in item.annotations
+                    if a.type == AnnotationType.bbox]
+
+                wider_annotation += '%s\n' % len(bboxes)
+                for bbox in bboxes:
+                    wider_bb = ' '.join('%d' % p for p in bbox.get_bbox())
+                    wider_annotation += '%s ' % wider_bb
+                    if bbox.attributes:
+                        for attr in WiderFacePath.BBOX_ATTRIBUTES:
+                            if bbox.attributes[attr]:
+                                wider_annotation += '%s ' % bbox.attributes[attr]
+                            else:
+                                wider_annotation += '0 '
+                    wider_annotation  += '\n'
+            annotation_path = osp.join(save_dir, WiderFacePath.ANNOTATIONS_DIR,
+                'wider_face_' + subset_name + '_bbx_gt.txt')
+            os.makedirs(osp.dirname(annotation_path), exist_ok=True)
+            with open(annotation_path, 'w') as f:
+                f.write(wider_annotation)

docs/user_manual.md

@@ -91,6 +91,9 @@ List of supported formats:
 - TF Detection API (`bboxes`, `masks`)
   - Format specifications: [bboxes](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/using_your_own_dataset.md), [masks](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/instance_segmentation.md)
   - [Dataset example](../tests/assets/tf_detection_api_dataset)
+- WIDER Face (`bboxes`)
+  - [Format specification](http://shuoyang1213.me/WIDERFACE/)
+  - [Dataset example](../tests/assets/wider_dataset)
 - MOT sequences
   - [Format specification](https://arxiv.org/pdf/1906.04567.pdf)
   - [Dataset example](../tests/assets/mot_dataset)

tests/assets/widerface_dataset/wider_face_split/wider_face_train_bbx_gt.txt

@@ -0,0 +1,7 @@
+0--Parade/0_Parade_image_01.jpg
+1
+1 2 2 2 0 0 0 0 0 0 
+1--Handshaking/1_Handshaking_image_02.jpg
+2
+1 1 2 2 0 0 1 0 0 0 
+5 1 2 2 0 0 1 0 0 0 

tests/assets/widerface_dataset/wider_face_split/wider_face_val_bbx_gt.txt

@@ -0,0 +1,5 @@
+0--Parade/0_Parade_image_03.jpg
+3
+0 0 1 1 2 0 0 0 2 0 
+3 2 1 2 0 0 0 1 0 0 
+5 6 1 1 2 0 0 0 2 0 

tests/test_widerface_format.py

@@ -0,0 +1,122 @@
+import os.path as osp
+from unittest import TestCase
+
+import numpy as np
+from datumaro.components.extractor import Bbox, DatasetItem
+from datumaro.components.project import Dataset, Project
+from datumaro.plugins.widerface_format import WiderFaceConverter, WiderFaceImporter
+from datumaro.util.test_utils import TestDir, compare_datasets
+
+
+class WiderFaceFormatTest(TestCase):
+    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, 1, 2, 3, attributes = {
+                        'blur': 2, 'expression': 0, 'illumination': 0,
+                        'occlusion': 0, 'pose': 2, 'invalid': 0}),
+                ]
+            ),
+            DatasetItem(id='2', subset='train', image=np.ones((10, 10, 3)),
+                annotations=[
+                    Bbox(0, 2, 4, 2, attributes = {
+                        'blur': 2, 'expression': 0, 'illumination': 1,
+                        'occlusion': 0, 'pose': 1, 'invalid': 0}),
+                    Bbox(3, 3, 2, 3, attributes = {
+                        'blur': 0, 'expression': 1, 'illumination': 0,
+                        'occlusion': 0, 'pose': 2, 'invalid': 0}),
+                    Bbox(2, 1, 2, 3, attributes = {
+                        'blur': 2, 'expression': 0, 'illumination': 0,
+                        'occlusion': 0, 'pose': 0, 'invalid': 1}),
+                ]
+            ),
+
+            DatasetItem(id='3', subset='val', image=np.ones((8, 8, 3)),
+                annotations=[
+                    Bbox(0, 1, 5, 2, attributes = {
+                        'blur': 2, 'expression': 1, 'illumination': 0,
+                        'occlusion': 0, 'pose': 1, 'invalid': 0}),
+                    Bbox(0, 2, 3, 2),
+                    Bbox(0, 2, 4, 2),
+                    Bbox(0, 7, 3, 2, attributes = {
+                        'blur': 2, 'expression': 1, 'illumination': 0,
+                        'occlusion': 0, 'pose': 1, 'invalid': 0}),
+                ]
+            ),
+
+            DatasetItem(id='4', subset='val', image=np.ones((8, 8, 3))),
+        ])
+
+        with TestDir() as test_dir:
+            WiderFaceConverter.convert(source_dataset, test_dir, save_images=True)
+            parsed_dataset = WiderFaceImporter()(test_dir).make_dataset()
+
+            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)),
+                annotations=[
+                    Bbox(0, 2, 4, 2),
+                    Bbox(0, 1, 2, 3, attributes = {
+                        'blur': 2, 'expression': 0, 'illumination': 0,
+                        'occlusion': 0, 'pose': 2, 'invalid': 0}),
+                ]
+            ),
+        ])
+
+        with TestDir() as test_dir:
+            WiderFaceConverter.convert(source_dataset, test_dir, save_images=True)
+            parsed_dataset = WiderFaceImporter()(test_dir).make_dataset()
+
+            compare_datasets(self, source_dataset, parsed_dataset)
+
+DUMMY_DATASET_DIR = osp.join(osp.dirname(__file__), 'assets', 'widerface_dataset')
+
+class WiderFaceImporterTest(TestCase):
+    def test_can_detect(self):
+        self.assertTrue(WiderFaceImporter.detect(DUMMY_DATASET_DIR))
+
+    def test_can_import(self):
+        expected_dataset = Dataset.from_iterable([
+            DatasetItem(id='0--Parade/0_Parade_image_01', subset='train',
+                image=np.ones((10, 15, 3)),
+                annotations=[
+                    Bbox(1, 2, 2, 2, attributes = {
+                        'blur': 0, 'expression': 0, 'illumination': 0,
+                        'occlusion': 0, 'pose': 0, 'invalid': 0}),
+                ]
+            ),
+            DatasetItem(id='1--Handshaking/1_Handshaking_image_02', subset='train',
+                image=np.ones((10, 15, 3)),
+                annotations=[
+                    Bbox(1, 1, 2, 2, attributes = {
+                        'blur': 0, 'expression': 0, 'illumination': 1,
+                        'occlusion': 0, 'pose': 0, 'invalid': 0}),
+                    Bbox(5, 1, 2, 2, attributes = {
+                        'blur': 0, 'expression': 0, 'illumination': 1,
+                        'occlusion': 0, 'pose': 0, 'invalid': 0}),
+                ]
+            ),
+            DatasetItem(id='0--Parade/0_Parade_image_03', subset='val',
+                image=np.ones((10, 15, 3)),
+                annotations=[
+                    Bbox(0, 0, 1, 1, attributes = {
+                        'blur': 2, 'expression': 0, 'illumination': 0,
+                        'occlusion': 0, 'pose': 2, 'invalid': 0}),
+                    Bbox(3, 2, 1, 2, attributes = {
+                        'blur': 0, 'expression': 0, 'illumination': 0,
+                        'occlusion': 1, 'pose': 0, 'invalid': 0}),
+                    Bbox(5, 6, 1, 1, attributes = {
+                        'blur': 2, 'expression': 0, 'illumination': 0,
+                        'occlusion': 0, 'pose': 2, 'invalid': 0}),
+                ]
+            ),
+        ])
+
+        dataset = Project.import_from(DUMMY_DATASET_DIR, 'wider_face') \
+            .make_dataset()
+
+        compare_datasets(self, expected_dataset, dataset)