add ImagenetFullBboxDataset

chainercv/datasets/__init__.py

@@ -19,6 +19,15 @@
 from chainercv.datasets.cub.cub_utils import cub_label_names  # NOQA
 from chainercv.datasets.directory_parsing_label_dataset import directory_parsing_label_names  # NOQA
 from chainercv.datasets.directory_parsing_label_dataset import DirectoryParsingLabelDataset  # NOQA
+from chainercv.datasets.imagenet.imagenet_det_bbox_dataset import ImagenetDetBboxDataset  # NOQA
+from chainercv.datasets.imagenet.imagenet_full_bbox_dataset import ImagenetFullBboxDataset  # NOQA
+from chainercv.datasets.imagenet.imagenet_loc_bbox_dataset import ImagenetLocBboxDataset  # NOQA
+from chainercv.datasets.imagenet.imagenet_utils import get_imagenet_full_bbox_label_names  # NOQA
+from chainercv.datasets.imagenet.imagenet_utils import imagenet_det_bbox_label_names  # NOQA
+from chainercv.datasets.imagenet.imagenet_utils import imagenet_det_synset_ids  # NOQA
+from chainercv.datasets.imagenet.imagenet_utils import imagenet_full_bbox_synset_ids  # NOQA
+from chainercv.datasets.imagenet.imagenet_utils import imagenet_loc_bbox_label_names  # NOQA
+from chainercv.datasets.imagenet.imagenet_utils import imagenet_loc_synset_ids  # NOQA
 from chainercv.datasets.mixup_soft_label_dataset import MixUpSoftLabelDataset  # 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

chainercv/datasets/imagenet/imagenet_det_bbox_dataset.py

@@ -0,0 +1,172 @@
+import numpy as np
+import os
+
+from chainer.dataset import download
+
+from chainercv.chainer_experimental.datasets.sliceable import GetterDataset
+from chainercv.datasets.imagenet.imagenet_utils import get_ilsvrc_devkit
+from chainercv.datasets.imagenet.imagenet_utils import imagenet_det_synset_ids
+from chainercv.datasets.voc.voc_utils import parse_voc_bbox_annotation
+from chainercv.utils import read_image
+
+
+class ImagenetDetBboxDataset(GetterDataset):
+
+    """ILSVRC ImageNet detection dataset.
+
+    The data is distributed on the `official Kaggle page`_.
+
+    .. _`official Kaggle page`: https://www.kaggle.com/c/
+        imagenet-object-detection-challenge
+
+    Please refer to the readme of ILSVRC2014 dev kit for a comprehensive
+    documentation. Note that the detection part of ILSVRC has not changed since
+    2014. An overview of annotation process is described in the `paper`_.
+
+    .. _`paper`: http://ai.stanford.edu/~olga/papers/chi2014-MultiLabel.pdf
+
+    Every image in the training set has one or more image-level labels.
+    The image-level labels determine the full presence, partial presence or
+    absence of one or more object categories.
+    Bounding boxes are provided around instances of the present categories.
+
+    Args:
+        data_dir (string): Path to the root of the training data. If this is
+            :obj:`auto`,
+            :obj:`$CHAINER_DATASET_ROOT/pfnet/chainercv/imagenet` is used.
+        split ({'train', 'val', 'val1', 'val2'}): Selects a split of the
+            dataset.
+        year ({'2013', '2014'}): Use a dataset prepared for a challenge
+            held in :obj:`year`. The default value is :obj:`2014`.
+        return_img_label (bool): If :obj:`True`, this dataset returns
+            image-wise labels. This consists of two arrays:
+            :obj:`img_label` and :obj:`img_label_type`.
+        use_val_blacklist (bool): If :obj:`False`, images that are
+            included in the blacklist are avoided when
+            the split is :obj:`val`. The default value is :obj:`False`.
+
+    This dataset returns the following data.
+
+    .. csv-table::
+        :header: name, shape, dtype, format
+
+        :obj:`img`, ":math:`(3, H, W)`", :obj:`float32`, \
+        "RGB, :math:`[0, 255]`"
+        :obj:`bbox`, ":math:`(R, 4)`", :obj:`float32`, \
+        ":math:`(y_{min}, x_{min}, y_{max}, x_{max})`"
+        :obj:`label`, ":math:`(R,)`", :obj:`int32`, \
+        ":math:`[0, \#fg\_class - 1]`"
+        :obj:`img_label` [#imagenet_det_1]_, ":math:`(M,)`", :obj:`int32`, \
+        ":math:`[0, \#fg\_class - 1]`"
+        :obj:`img_label_type` [#imagenet_det_1]_ [#imagenet_det_2]_, \
+            ":math:`(M,)`", :obj:`int32`, ":math:`[-1, 1]`"
+
+    .. [#imagenet_det_1] available
+        if :obj:`return_img_label = True`.
+    .. [#imagenet_det_2] :obj:`-1` means absent. :obj:`1` means present.
+        :obj:`0` means partially present. When a category is partially
+        present, the image contains at least one
+        instance of X, but not all instances of X may be annotated with
+        bounding boxes.
+
+    """
+
+    def __init__(self, data_dir='auto', split='train', year='2014',
+                 return_img_label=False, use_val_blacklist=False):
+        super(ImagenetDetBboxDataset, self).__init__()
+        if data_dir == 'auto':
+            data_dir = download.get_dataset_directory(
+                'pfnet/chainercv/imagenet')
+            get_ilsvrc_devkit()
+        val_blacklist_path = os.path.join(
+            data_dir, 'ILSVRC2014_devkit/data/',
+            'ILSVRC2014_det_validation_blacklist.txt')
+
+        if year not in ('2013', '2014'):
+            raise ValueError('\'year\' has to be either '
+                             '\'2013\' or \'2014\'.')
+        self.base_dir = os.path.join(data_dir, 'ILSVRC')
+        imageset_dir = os.path.join(self.base_dir, 'ImageSets/DET')
+
+        if split == 'train':
+            img_labels = {}
+            for lb in range(0, 200):
+                with open(os.path.join(
+                        imageset_dir, 'train_{}.txt'.format(lb + 1))) as f:
+                    for l in f:
+                        id_ = l.split()[0]
+                        if 'ILSVRC2014' in id_ and year != '2014':
+                            continue
+
+                        anno_type = l.split()[1]
+                        if id_ not in img_labels:
+                            img_labels[id_] = []
+                        img_labels[id_].append((lb, int(anno_type)))
+                self.img_labels = img_labels
+                self.ids = list(img_labels.keys())
+                self.split_type = 'train'
+        else:
+            if return_img_label:
+                raise ValueError('split has to be \'train\' when '
+                                 'return_img_label is True')
+            if use_val_blacklist:
+                blacklist_ids = []
+            else:
+                ids = []
+                with open(os.path.join(
+                        imageset_dir, 'val.txt'.format(split))) as f:
+                    for l in f:
+                        id_ = l.split()[0]
+                        ids.append(id_)
+                blacklist_ids = []
+                with open(val_blacklist_path) as f:
+                    for l in f:
+                        index = int(l.split()[0])
+                        blacklist_ids.append(ids[index])
+
+            ids = []
+            with open(os.path.join(
+                    imageset_dir, '{}.txt'.format(split))) as f:
+                for l in f:
+                    id_ = l.split()[0]
+                    if id_ not in blacklist_ids:
+                        ids.append(id_)
+                self.ids = ids
+            self.split_type = 'val'
+
+        self.add_getter('img', self._get_image)
+        self.add_getter(('bbox', 'label'), self._get_inst_anno)
+        if return_img_label:
+            self.add_getter(
+                ('img_label', 'img_label_type'), self._get_img_label)
+
+    def __len__(self):
+        return len(self.ids)
+
+    def _get_image(self, i):
+        img_path = os.path.join(
+            self.base_dir, 'Data/DET', self.split_type,
+            self.ids[i] + '.JPEG')
+        img = read_image(img_path, color=True)
+        return img
+
+    def _get_inst_anno(self, i):
+        if 'extra' not in self.ids[i]:
+            anno_path = os.path.join(
+                self.base_dir, 'Annotations/DET', self.split_type,
+                self.ids[i] + '.xml')
+            bbox, label, _ = parse_voc_bbox_annotation(
+                anno_path, imagenet_det_synset_ids,
+                skip_names_not_in_label_names=True)
+        else:
+            bbox = np.zeros((0, 4), dtype=np.float32)
+            label = np.zeros((0,), dtype=np.int32)
+        return bbox, label
+
+    def _get_img_label(self, i):
+        img_label = np.array([val[0] for val in self.img_labels[self.ids[i]]],
+                             dtype=np.int32)
+        img_label_type = np.array(
+            [val[1] for val in self.img_labels[self.ids[i]]],
+            dtype=np.int32)
+        return img_label, img_label_type

chainercv/datasets/imagenet/imagenet_full_bbox_dataset.py

@@ -0,0 +1,125 @@
+import numpy as np
+import os
+
+from chainer.dataset import download
+
+from chainercv.chainer_experimental.datasets.sliceable import GetterDataset
+from chainercv.datasets.imagenet.imagenet_utils import \
+    imagenet_full_bbox_synset_ids
+from chainercv.datasets.voc.voc_utils import parse_voc_bbox_annotation
+from chainercv.utils import read_image
+
+
+class ImagenetFullBboxDataset(GetterDataset):
+
+    """ImageNet with bounding box annotation.
+
+    There are 3627 categories in this dataset.
+
+    Readable label names can be found by
+    :obj:`chainercv.datasets.get_imagenet_full_bbox_label_names`.
+
+    The data needs to be downloaded from the official page.
+    There are four steps to prepare data.
+
+    1. Download annotations from \
+        http://image-net.org/Annotation/Annotation.tar.gz.
+    2. Expand it under :obj:`DATA_DIR/Full/Annotation` with the following \
+        command.
+
+    .. code::
+
+        find -name  "*.tar.gz" | while read NAME ; \\
+            do tar -xvf "${NAME%.tar.gz}.tar.gz" ; done
+
+    3. Download images using ImageNet API for all synset ids in \
+        :obj:`imagenet_full_bbox_synset_ids`. Images for each synset \
+        can be downloaded by a command like below. You need to register \
+        ImageNet website to use this API.
+
+    .. code::
+
+        wget -O Data/<id>.tar "http://www.image-net.org/download/synset? \\
+            wnid=<id>&username=<username>&accesskey=<key>&release=latest& \\
+            src=stanford
+
+    4. Expand images under :obj:`DATA_DIR/Full/Data` with the following \
+        command.
+
+    .. code::
+
+        find -name  "*.tar" | while read NAME ; do mkdir ${NAME%.tar}; \\
+            mv ${NAME%.tar}.tar ${NAME%.tar}; cd ${NAME%.tar}; \\
+            tar -xvf ${NAME%.tar}.tar; rm ${NAME%.tar}.tar ; cd ..; done
+
+    Args:
+        data_dir (string): Path to the root of the data. If this is
+            :obj:`auto`,
+            :obj:`$CHAINER_DATASET_ROOT/pfnet/chainercv/imagenet` is used.
+        return_img_label (bool): If :obj:`True`, this dataset returns
+            image-wise labels.
+
+    This dataset returns the following data.
+
+    .. csv-table::
+        :header: name, shape, dtype, format
+
+        :obj:`img`, ":math:`(3, H, W)`", :obj:`float32`, \
+        "RGB, :math:`[0, 255]`"
+        :obj:`bbox`, ":math:`(R, 4)`", :obj:`float32`, \
+        ":math:`(y_{min}, x_{min}, y_{max}, x_{max})`"
+        :obj:`label`, ":math:`(R,)`", :obj:`int32`, \
+        ":math:`[0, \#fg\_class - 1]`"
+        :obj:`img_label` [#imagenet_full_1]_, ":math:`()`", :obj:`int32`, \
+        ":math:`[0, \#fg\_class - 1]`"
+
+    .. [#imagenet_full_1] available
+        if :obj:`return_img_label = True`.
+
+    """
+
+    def __init__(self, data_dir='auto',
+                 return_img_label=False):
+        super(ImagenetFullBboxDataset, self).__init__()
+        if data_dir == 'auto':
+            data_dir = download.get_dataset_directory(
+                'pfnet/chainercv/imagenet')
+        self.base_dir = os.path.join(data_dir, 'Full')
+
+        self.paths = []
+        self.cls_names = []
+        self.img_paths = []
+        image_dir = os.path.join(self.base_dir, 'Annotation')
+        for cls_name in sorted(os.listdir(image_dir)):
+            cls_dir = os.path.join(image_dir, cls_name)
+            for name in sorted(os.listdir(cls_dir)):
+                img_path = os.path.join(
+                    self.base_dir, 'Data', cls_name, name[:-4] + '.JPEG')
+                if os.path.exists(img_path):
+                    self.paths.append(os.path.join(cls_dir, name))
+                    self.img_paths.append(img_path)
+                    self.cls_names.append(cls_name)
+
+        self.add_getter('img', self._get_image)
+        self.add_getter(('bbox', 'label'), self._get_inst_anno)
+        self.add_getter('img_label', self._get_img_label)
+        if not return_img_label:
+            self.keys = ('img', 'bbox', 'label')
+
+    def __len__(self):
+        return len(self.paths)
+
+    def _get_image(self, i):
+        img = read_image(self.img_paths[i], color=True)
+        return img
+
+    def _get_inst_anno(self, i):
+        bbox, label, _ = parse_voc_bbox_annotation(
+            self.paths[i], imagenet_full_bbox_synset_ids,
+            skip_names_not_in_label_names=False)
+        return bbox, label
+
+    def _get_img_label(self, i):
+        label = imagenet_full_bbox_synset_ids.index(
+            self.cls_names[i])
+        return np.array([label], dtype=np.int32)