def __iter__(self):
return iter([
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,
'truncated': True,
'difficult': False,
'occluded': False,
}),
Bbox(2,
3,
1,
1,
label=self._label(VOC.VocBodyPart(1).name),
group=1),
Bbox(5,
4,
3,
2,
label=self._label(VOC.VocBodyPart(2).name),
group=1),
]),
])
def __iter__(self):
return iter([
DatasetItem(id='2007_000001', subset='train',
annotations=[
Bbox(4, 5, 2, 2, label=self._label('person'),
attributes={
'truncated': False,
'difficult': False,
'occluded': False,
**{
a.name: a.value % 2 == 1
for a in VOC.VocAction
}
},
id=2, group=2,
# TODO: Actions should be excluded
# as soon as correct merge is implemented
),
Bbox(5.5, 6, 2, 2, label=self._label(
VOC.VocBodyPart(1).name),
group=2
)
]
),
DatasetItem(id='2007_000002', subset='test')
])
def __iter__(self):
return iter([
DatasetItem(
id='2007_000001',
subset='train',
image=Image(path='2007_000001.jpg', size=(10, 20)),
annotations=[
Label(self._label(l.name))
for l in VOC.VocLabel if l.value % 2 == 1
] + [
Bbox(
1,
2,
2,
2,
label=self._label('cat'),
attributes={
'pose': VOC.VocPose(1).name,
'truncated': True,
'difficult': False,
'occluded': False,
},
id=1,
group=1,
),
Bbox(
4,
5,
2,
2,
label=self._label('person'),
attributes={
'truncated': False,
'difficult': False,
'occluded': False,
**{
a.name: a.value % 2 == 1
for a in VOC.VocAction
}
},
id=2,
group=2,
),
Bbox(5.5,
6,
2,
2,
label=self._label(VOC.VocBodyPart(1).name),
group=2),
Mask(
image=np.ones([5, 10]),
label=self._label(VOC.VocLabel(2).name),
group=1,
),
]),
DatasetItem(id='2007_000002',
subset='test',
image=np.ones((10, 20, 3))),
])
def generate_dummy_voc(path):
cls_subsets_dir = osp.join(path, 'ImageSets', 'Main')
action_subsets_dir = osp.join(path, 'ImageSets', 'Action')
layout_subsets_dir = osp.join(path, 'ImageSets', 'Layout')
segm_subsets_dir = osp.join(path, 'ImageSets', 'Segmentation')
ann_dir = osp.join(path, 'Annotations')
img_dir = osp.join(path, 'JPEGImages')
segm_dir = osp.join(path, 'SegmentationClass')
inst_dir = osp.join(path, 'SegmentationObject')
os.makedirs(cls_subsets_dir)
os.makedirs(ann_dir)
os.makedirs(img_dir)
os.makedirs(segm_dir)
os.makedirs(inst_dir)
subsets = {
'train': ['2007_000001'],
'test': ['2007_000002'],
}
# Subsets
for subset_name, subset in subsets.items():
for item in subset:
with open(osp.join(cls_subsets_dir, subset_name + '.txt'), 'w') as f:
for item in subset:
f.write('%s\n' % item)
shutil.copytree(cls_subsets_dir, action_subsets_dir)
shutil.copytree(cls_subsets_dir, layout_subsets_dir)
shutil.copytree(cls_subsets_dir, segm_subsets_dir)
# Classification
subset_name = 'train'
subset = subsets[subset_name]
for label in VOC.VocLabel:
with open(osp.join(cls_subsets_dir, '%s_%s.txt' % \
(label.name, subset_name)), 'w') as f:
for item in subset:
presence = label.value % 2
f.write('%s %2d\n' % (item, 1 if presence else -1))
# Detection + Action + Layout
subset_name = 'train'
subset = subsets[subset_name]
for item in subset:
root_elem = ET.Element('annotation')
ET.SubElement(root_elem, 'folder').text = 'VOC' + item.split('_')[0]
ET.SubElement(root_elem, 'filename').text = item + '.jpg'
size_elem = ET.SubElement(root_elem, 'size')
ET.SubElement(size_elem, 'width').text = '10'
ET.SubElement(size_elem, 'height').text = '20'
ET.SubElement(size_elem, 'depth').text = '3'
ET.SubElement(root_elem, 'segmented').text = '1'
obj1_elem = ET.SubElement(root_elem, 'object')
ET.SubElement(obj1_elem, 'name').text = 'cat'
ET.SubElement(obj1_elem, 'pose').text = VOC.VocPose(1).name
ET.SubElement(obj1_elem, 'truncated').text = '1'
ET.SubElement(obj1_elem, 'difficult').text = '0'
obj1bb_elem = ET.SubElement(obj1_elem, 'bndbox')
ET.SubElement(obj1bb_elem, 'xmin').text = '1'
ET.SubElement(obj1bb_elem, 'ymin').text = '2'
ET.SubElement(obj1bb_elem, 'xmax').text = '3'
ET.SubElement(obj1bb_elem, 'ymax').text = '4'
obj2_elem = ET.SubElement(root_elem, 'object')
ET.SubElement(obj2_elem, 'name').text = 'person'
obj2bb_elem = ET.SubElement(obj2_elem, 'bndbox')
ET.SubElement(obj2bb_elem, 'xmin').text = '4'
ET.SubElement(obj2bb_elem, 'ymin').text = '5'
ET.SubElement(obj2bb_elem, 'xmax').text = '6'
ET.SubElement(obj2bb_elem, 'ymax').text = '7'
obj2head_elem = ET.SubElement(obj2_elem, 'part')
ET.SubElement(obj2head_elem, 'name').text = VOC.VocBodyPart(1).name
obj2headbb_elem = ET.SubElement(obj2head_elem, 'bndbox')
ET.SubElement(obj2headbb_elem, 'xmin').text = '5.5'
ET.SubElement(obj2headbb_elem, 'ymin').text = '6'
ET.SubElement(obj2headbb_elem, 'xmax').text = '7.5'
ET.SubElement(obj2headbb_elem, 'ymax').text = '8'
obj2act_elem = ET.SubElement(obj2_elem, 'actions')
for act in VOC.VocAction:
ET.SubElement(obj2act_elem, act.name).text = '%s' % (act.value % 2)
with open(osp.join(ann_dir, item + '.xml'), 'w') as f:
f.write(ET.tostring(root_elem, encoding='unicode'))
# Segmentation + Instances
subset_name = 'train'
subset = subsets[subset_name]
for item in subset:
save_image(osp.join(segm_dir, item + '.png'),
np.tile(VOC.VocColormap[2][::-1], (5, 10, 1))
)
save_image(osp.join(inst_dir, item + '.png'),
np.tile(1, (5, 10, 1)))
# Test images
subset_name = 'test'
subset = subsets[subset_name]
for item in subset:
save_image(osp.join(img_dir, item + '.jpg'),
np.ones([10, 20, 3]))
return subsets
