Add VOC and SBD instance segmentation dataset

README.md

@@ -40,6 +40,7 @@ For additional features
 
 + Matplotlib
 + OpenCV
++ SciPy
 
 Environments under Python 2.7.12 and 3.6.0 are tested.
 

chainercv/datasets/__init__.py

@@ -17,11 +17,15 @@
 from chainercv.datasets.directory_parsing_label_dataset import DirectoryParsingLabelDataset  # NOQA
 from chainercv.datasets.online_products.online_products_dataset import online_products_super_label_names  # NOQA
 from chainercv.datasets.online_products.online_products_dataset import OnlineProductsDataset  # NOQA
+from chainercv.datasets.sbd.sbd_instance_segmentation_dataset import SBDInstanceSegmentationDataset  # NOQA
+from chainercv.datasets.sbd.sbd_utils import sbd_instance_segmentation_label_names  # NOQA
 from chainercv.datasets.siamese_dataset import SiameseDataset  # NOQA
 from chainercv.datasets.transform_dataset import TransformDataset  # NOQA
 from chainercv.datasets.voc.voc_bbox_dataset import VOCBboxDataset  # NOQA
+from chainercv.datasets.voc.voc_instance_segmentation_dataset import VOCInstanceSegmentationDataset  # NOQA
 from chainercv.datasets.voc.voc_semantic_segmentation_dataset import VOCSemanticSegmentationDataset  # NOQA
 from chainercv.datasets.voc.voc_utils import voc_bbox_label_names  # NOQA
+from chainercv.datasets.voc.voc_utils import voc_instance_segmentation_label_names  # NOQA
 from chainercv.datasets.voc.voc_utils import voc_semantic_segmentation_ignore_label_color  # NOQA
 from chainercv.datasets.voc.voc_utils import voc_semantic_segmentation_label_colors  # NOQA
 from chainercv.datasets.voc.voc_utils import voc_semantic_segmentation_label_names  # NOQA

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

@@ -0,0 +1,103 @@
+import numpy as np
+import os
+import warnings
+
+import chainer
+
+from chainercv.datasets.sbd import sbd_utils
+from chainercv.datasets.voc import voc_utils
+from chainercv.utils import read_image
+
+try:
+    import scipy
+    _available = True
+except ImportError:
+    _available = False
+
+
+def _check_available():
+    if not _available:
+        warnings.warn(
+            'SciPy is not installed in your environment,',
+            'so the dataset cannot be loaded.'
+            'Please install SciPy to load dataset.\n\n'
+            '$ pip install scipy')
+
+
+class SBDInstanceSegmentationDataset(chainer.dataset.DatasetMixin):
+
+    """Instance segmentation dataset for Semantic Boundaries Dataset `SBD`_.
+
+    The class name of the label :math:`l` is :math:`l` th element of
+    :obj:`chainercv.datasets.sbd_instance_segmentation_label_names`.
+
+    .. _`SBD`: http://home.bharathh.info/pubs/codes/SBD/download.html
+
+    Args:
+        data_dir (string): Path to the root of the training data. If this is
+            :obj:`auto`, this class will automatically download data for you
+            under :obj:`$CHAINER_DATASET_ROOT/pfnet/chainercv/sbd`.
+        split ({'train', 'val', 'trainval'}): Select a split of the dataset.
+
+    """
+
+    def __init__(self, data_dir='auto', split='train'):
+        _check_available()
+
+        if split not in ['train', 'trainval', 'val']:
+            raise ValueError(
+                'please pick split from \'train\', \'trainval\', \'val\'')
+
+        if data_dir == 'auto':
+            data_dir = sbd_utils.get_sbd()
+
+        id_list_file = os.path.join(
+            data_dir, '{}_voc2012.txt'.format(split))
+        self.ids = [id_.strip() for id_ in open(id_list_file)]
+
+        self.data_dir = data_dir
+
+    def __len__(self):
+        return len(self.ids)
+
+    def get_example(self, i):
+        """Returns the i-th example.
+
+        Returns a color image, bounding boxes, masks and labels. The color
+        image is in CHW format.
+
+        Args:
+            i (int): The index of the example.
+
+        Returns:
+            A tuple of color image, bounding boxes, masks and labels whose
+            shapes are :math:`(3, H, W), (R, 4), (R, H, W), (R, )`
+            respectively.
+            :math:`H` and :math:`W` are height and width of the images,
+            and :math:`R` is the number of objects in the image.
+            The dtype of the color image and the bounding boxes are
+            :obj:`numpy.float32`, that of the masks is :obj: `numpy.bool`,
+            and that of the labels is :obj:`numpy.int32`.
+
+        """
+        data_id = self.ids[i]
+        img_file = os.path.join(
+            self.data_dir, 'img', data_id + '.jpg')
+        img = read_image(img_file, color=True)
+        label_img, inst_img = self._load_label_inst(data_id)
+        bbox, mask, label = voc_utils.image_wise_to_instance_wise(
+            label_img, inst_img)
+        return img, bbox, mask, label
+
+    def _load_label_inst(self, data_id):
+        label_file = os.path.join(
+            self.data_dir, 'cls', data_id + '.mat')
+        inst_file = os.path.join(
+            self.data_dir, 'inst', data_id + '.mat')
+        label_anno = scipy.io.loadmat(label_file)
+        label_img = label_anno['GTcls']['Segmentation'][0][0].astype(np.int32)
+        inst_anno = scipy.io.loadmat(inst_file)
+        inst_img = inst_anno['GTinst']['Segmentation'][0][0].astype(np.int32)
+        inst_img[inst_img == 0] = -1
+        inst_img[inst_img == 255] = -1
+        return label_img, inst_img

chainercv/datasets/sbd/sbd_utils.py

@@ -0,0 +1,50 @@
+import os
+import six
+
+from chainer.dataset import download
+
+from chainercv.datasets.voc.voc_utils \
+    import voc_instance_segmentation_label_names
+from chainercv import utils
+
+root = 'pfnet/chainercv/sbd'
+url = 'http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/semantic_contours/benchmark.tgz'  # NOQA
+train_voc2012_url = 'http://home.bharathh.info/pubs/codes/SBD/train_noval.txt'
+
+
+def _generate_voc2012_txt(base_path):
+    with open(os.path.join(base_path, 'train.txt'), 'r') as f:
+        train_ids = f.read().split('\n')[:-1]
+    with open(os.path.join(base_path, 'val.txt'), 'r') as f:
+        val_ids = f.read().split('\n')[:-1]
+    with open(os.path.join(base_path, 'train_voc2012.txt'), 'r') as f:
+        train_voc2012_ids = f.read().split('\n')[:-1]
+    all_ids = list(set(train_ids + val_ids))
+    val_voc2012_ids = [i for i in all_ids if i not in train_voc2012_ids]
+
+    with open(os.path.join(base_path, 'val_voc2012.txt'), 'w') as f:
+        f.write('\n'.join(sorted(val_voc2012_ids)))
+    with open(os.path.join(base_path, 'trainval_voc2012.txt'), 'w') as f:
+        f.write('\n'.join(sorted(all_ids)))
+
+
+def get_sbd():
+    data_root = download.get_dataset_directory(root)
+    base_path = os.path.join(data_root, 'benchmark_RELEASE/dataset')
+
+    train_voc2012_file = os.path.join(base_path, 'train_voc2012.txt')
+    if os.path.exists(train_voc2012_file):
+        # skip downloading
+        return base_path
+
+    download_file_path = utils.cached_download(url)
+    ext = os.path.splitext(url)[1]
+    utils.extractall(download_file_path, data_root, ext)
+
+    six.moves.urllib.request.urlretrieve(train_voc2012_url, train_voc2012_file)
+    _generate_voc2012_txt(base_path)
+
+    return base_path
+
+
+sbd_instance_segmentation_label_names = voc_instance_segmentation_label_names

chainercv/datasets/voc/voc_instance_segmentation_dataset.py

@@ -0,0 +1,85 @@
+import numpy as np
+import os
+
+import chainer
+
+from chainercv.datasets.voc import voc_utils
+from chainercv.utils import read_image
+
+
+class VOCInstanceSegmentationDataset(chainer.dataset.DatasetMixin):
+
+    """Instance segmentation dataset for PASCAL `VOC2012`_.
+
+    The class name of the label :math:`l` is :math:`l` th element of
+    :obj:`chainercv.datasets.voc_instance_segmentation_label_names`.
+
+    .. _`VOC2012`: http://host.robots.ox.ac.uk/pascal/VOC/voc2012/
+
+    Args:
+        data_dir (string): Path to the root of the training data. If this is
+            :obj:`auto`, this class will automatically download data for you
+            under :obj:`$CHAINER_DATASET_ROOT/pfnet/chainercv/voc`.
+        split ({'train', 'val', 'trainval'}): Select a split of the dataset.
+
+    """
+
+    def __init__(self, data_dir='auto', split='train'):
+        if split not in ['train', 'trainval', 'val']:
+            raise ValueError(
+                'please pick split from \'train\', \'trainval\', \'val\'')
+
+        if data_dir == 'auto':
+            data_dir = voc_utils.get_voc('2012', split)
+
+        id_list_file = os.path.join(
+            data_dir, 'ImageSets/Segmentation/{0}.txt'.format(split))
+        self.ids = [id_.strip() for id_ in open(id_list_file)]
+
+        self.data_dir = data_dir
+
+    def __len__(self):
+        return len(self.ids)
+
+    def get_example(self, i):
+        """Returns the i-th example.
+
+        Returns a color image, bounding boxes, masks and labels. The color
+        image is in CHW format.
+
+        Args:
+            i (int): The index of the example.
+
+        Returns:
+            A tuple of color image, bounding boxes, masks and labels whose
+            shapes are :math:`(3, H, W), (R, 4), (R, H, W), (R, )`
+            respectively.
+            :math:`H` and :math:`W` are height and width of the images,
+            and :math:`R` is the number of objects in the image.
+            The dtype of the color image and the bounding boxes are
+            :obj:`numpy.float32`, that of the masks is :obj: `numpy.bool`,
+            and that of the labels is :obj:`numpy.int32`.
+
+       """
+        data_id = self.ids[i]
+        img_file = os.path.join(
+            self.data_dir, 'JPEGImages', data_id + '.jpg')
+        img = read_image(img_file, color=True)
+        label_img, inst_img = self._load_label_inst(data_id)
+        bbox, mask, label = voc_utils.image_wise_to_instance_wise(
+            label_img, inst_img)
+        return img, bbox, mask, label
+
+    def _load_label_inst(self, data_id):
+        label_file = os.path.join(
+            self.data_dir, 'SegmentationClass', data_id + '.png')
+        inst_file = os.path.join(
+            self.data_dir, 'SegmentationObject', data_id + '.png')
+        label_img = read_image(label_file, dtype=np.int32, color=False)
+        label_img = label_img[0]
+        label_img[label_img == 255] = -1
+        inst_img = read_image(inst_file, dtype=np.int32, color=False)
+        inst_img = inst_img[0]
+        inst_img[inst_img == 0] = -1
+        inst_img[inst_img == 255] = -1
+        return label_img, inst_img

chainercv/datasets/voc/voc_utils.py

@@ -1,3 +1,4 @@
+import numpy as np
 import os
 
 from chainer.dataset import download
@@ -37,6 +38,30 @@ def get_voc(year, split):
     return base_path
 
 
+def image_wise_to_instance_wise(label_img, inst_img):
+    bbox = []
+    mask = []
+    label = []
+    inst_ids = np.unique(inst_img)
+    for inst_id in inst_ids[inst_ids != -1]:
+        msk = inst_img == inst_id
+        lbl = np.unique(label_img[msk])[0] - 1
+
+        assert inst_id != -1
+        assert lbl != -1
+
+        where = np.argwhere(msk)
+        (y_min, x_min), (y_max, x_max) = where.min(0), where.max(0) + 1
+
+        bbox.append((y_min, x_min, y_max, x_max))
+        mask.append(msk)
+        label.append(lbl)
+    bbox = np.array(bbox).astype(np.float32)
+    mask = np.array(mask).astype(np.bool)
+    label = np.array(label).astype(np.int32)
+    return bbox, mask, label
+
+
 voc_bbox_label_names = (
     'aeroplane',
     'bicycle',
@@ -62,6 +87,8 @@ def get_voc(year, split):
 voc_semantic_segmentation_label_names = (('background',) +
                                          voc_bbox_label_names)
 
+voc_instance_segmentation_label_names = voc_bbox_label_names
+
 # these colors are used in the original MATLAB tools
 voc_semantic_segmentation_label_colors = (
     (0, 0, 0),

chainercv/utils/__init__.py

@@ -13,6 +13,7 @@
 from chainercv.utils.testing import assert_is_bbox_dataset  # NOQA
 from chainercv.utils.testing import assert_is_detection_link  # NOQA
 from chainercv.utils.testing import assert_is_image  # NOQA
+from chainercv.utils.testing import assert_is_instance_segmentation_dataset  # NOQA
 from chainercv.utils.testing import assert_is_label_dataset  # NOQA
 from chainercv.utils.testing import assert_is_point  # NOQA
 from chainercv.utils.testing import assert_is_semantic_segmentation_dataset  # NOQA

chainercv/utils/testing/__init__.py

@@ -2,6 +2,7 @@
 from chainercv.utils.testing.assertions import assert_is_bbox_dataset  # NOQA
 from chainercv.utils.testing.assertions import assert_is_detection_link  # NOQA
 from chainercv.utils.testing.assertions import assert_is_image  # NOQA
+from chainercv.utils.testing.assertions import assert_is_instance_segmentation_dataset  # NOQA
 from chainercv.utils.testing.assertions import assert_is_label_dataset  # NOQA
 from chainercv.utils.testing.assertions import assert_is_point  # NOQA
 from chainercv.utils.testing.assertions import assert_is_semantic_segmentation_dataset  # NOQA

chainercv/utils/testing/assertions/__init__.py

@@ -2,6 +2,7 @@
 from chainercv.utils.testing.assertions.assert_is_bbox_dataset import assert_is_bbox_dataset  # NOQA
 from chainercv.utils.testing.assertions.assert_is_detection_link import assert_is_detection_link  # NOQA
 from chainercv.utils.testing.assertions.assert_is_image import assert_is_image  # NOQA
+from chainercv.utils.testing.assertions.assert_is_instance_segmentation_dataset import assert_is_instance_segmentation_dataset  # NOQA
 from chainercv.utils.testing.assertions.assert_is_label_dataset import assert_is_label_dataset  # NOQA
 from chainercv.utils.testing.assertions.assert_is_point import assert_is_point  # NOQA
 from chainercv.utils.testing.assertions.assert_is_semantic_segmentation_dataset import assert_is_semantic_segmentation_dataset  # NOQA

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

@@ -0,0 +1,64 @@
+import numpy as np
+import six
+
+from chainercv.utils.testing.assertions.assert_is_bbox import assert_is_bbox
+from chainercv.utils.testing.assertions.assert_is_image import assert_is_image
+
+
+def assert_is_instance_segmentation_dataset(
+        dataset, n_fg_class, n_example=None
+):
+    """Checks if a dataset satisfies instance segmentation dataset APIs.
+
+    This function checks if a given dataset satisfies instance segmentation
+    dataset APIs or not.
+    If the dataset does not satifiy the APIs, this function raises an
+    :class:`AssertionError`.
+
+    Args:
+        dataset: A dataset to be checked.
+        n_fg_class (int): The number of foreground classes.
+        n_example (int): The number of examples to be checked.
+            If this argument is specified, this function picks
+            examples ramdomly and checks them. Otherwise,
+            this function checks all examples.
+
+    """
+
+    assert len(dataset) > 0, 'The length of dataset must be greater than zero.'
+
+    if n_example:
+        for _ in six.moves.range(n_example):
+            i = np.random.randint(0, len(dataset))
+            _check_example(dataset[i], n_fg_class)
+    else:
+        for i in six.moves.range(len(dataset)):
+            _check_example(dataset[i], n_fg_class)
+
+
+def _check_example(example, n_fg_class):
+    assert len(example) >= 4, \
+        'Each example must have at least four elements:' \
+        'img, bbox, mask and label.'
+
+    img, bbox, mask, label = example[:4]
+
+    assert_is_image(img, color=True)
+    _, H, W = img.shape
+    assert_is_bbox(bbox, size=(H, W))
+    R = bbox.shape[0]
+
+    assert isinstance(mask, np.ndarray), \
+        'mask must be a numpy.ndarray.'
+    assert isinstance(label, np.ndarray), \
+        'label must be a numpy.ndarray.'
+    assert mask.dtype == np.bool, \
+        'The type of mask must be bool'
+    assert label.dtype == np.int32, \
+        'The type of label must be numpy.int32.'
+    assert mask.shape == (R, H, W), \
+        'The shape of mask must be (R, H, W).'
+    assert label.shape == (R,), \
+        'The shape of label must be (R, ).'
+    assert label.min() >= 0 and label.max() < n_fg_class, \
+        'The value of label must be in [0, n_fg_class - 1].'