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, label=2),
Bbox(0,
1,
2,
3,
label=1,
attributes={
'blur': '2',
'expression': '0',
'illumination': '0',
'occluded': '0',
'pose': '2',
'invalid': '0'
}),
]),
],
categories={
AnnotationType.label:
LabelCategories.from_iterable('label_' + str(i)
for i in range(3)),
})
with TestDir() as test_dir:
WiderFaceConverter.convert(source_dataset,
test_dir,
save_images=True)
parsed_dataset = Dataset.import_from(test_dir, 'wider_face')
compare_datasets(self, source_dataset, parsed_dataset)
def __iter__(self):
yield DatasetItem(id=1,
annotations=[
Bbox(2,
3,
4,
5,
label=self._label('label_1'),
id=1,
group=1,
attributes={
'truncated': False,
'difficult': False,
'occluded': False,
}),
Bbox(1,
2,
3,
4,
label=self._label('label_2'),
id=2,
group=2,
attributes={
'truncated': False,
'difficult': False,
'occluded': False,
}),
])
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={
'truncated': True,
'difficult': False,
'occluded': False,
# no attributes here in the label categories
}
),
Bbox(5, 4, 3, 2, label=self._label('person'),
id=2, group=2, attributes={
'truncated': True,
'difficult': False,
'occluded': False,
VOC.VocAction(1).name: True,
VOC.VocAction(2).name: True,
**{
a.name: False for a in VOC.VocAction
if a.value not in {1, 2}
}
}
),
]),
])
def __iter__(self):
return iter([
DatasetItem(id=1,
subset='a',
annotations=[
Bbox(2,
3,
4,
5,
label=2,
attributes={'occluded': True}),
Bbox(
2,
3,
4,
5,
label=3,
attributes={'truncated': True},
),
]),
DatasetItem(id=2,
subset='b',
annotations=[
Bbox(
5,
4,
6,
5,
label=3,
attributes={'difficult': True},
),
]),
])
def test_can_find_bbox_with_wrong_label(self):
detections = 3
class_count = 2
item1 = DatasetItem(id=1, annotations=[
Bbox(i * 10, 10, 10, 10, label=i)
for i in range(detections)
])
item2 = DatasetItem(id=2, annotations=[
Bbox(i * 10, 10, 10, 10, label=(i + 1) % class_count)
for i in range(detections)
])
iou_thresh = 0.5
comp = DistanceComparator(iou_threshold=iou_thresh)
result = comp.match_boxes(item1, item2)
matches, mispred, a_greater, b_greater = result
self.assertEqual(len(item1.annotations), len(mispred))
self.assertEqual(0, len(a_greater))
self.assertEqual(0, len(b_greater))
self.assertEqual(0, len(matches))
for a_bbox, b_bbox in mispred:
self.assertLess(iou_thresh, a_bbox.iou(b_bbox))
self.assertEqual((a_bbox.label + 1) % class_count, b_bbox.label)
def __iter__(self):
yield DatasetItem(id=1,
annotations=[
Bbox(2,
3,
4,
5,
label=self._label(
VOC.VocLabel(1).name),
id=1,
group=1,
attributes={
'truncated': False,
'difficult': False,
'occluded': False,
}),
Bbox(1,
2,
3,
4,
label=self._label('non_voc_label'),
id=2,
group=2,
attributes={
'truncated': False,
'difficult': False,
'occluded': False,
}),
])
def __iter__(self):
return iter([
DatasetItem(id='a/1', subset='a', annotations=[
Bbox(2, 3, 4, 5, label=2, id=1, group=1,
attributes={
'truncated': False,
'difficult': False,
'occluded': True,
}
),
Bbox(2, 3, 4, 5, label=3, id=2, group=2,
attributes={
'truncated': True,
'difficult': False,
'occluded': False,
},
),
]),
DatasetItem(id=2, subset='b', annotations=[
Bbox(5, 4, 6, 5, label=3, id=1, group=1,
attributes={
'truncated': False,
'difficult': True,
'occluded': False,
},
),
]),
])
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='a/b/1',
subset='a',
annotations=[
Bbox(2,
3,
4,
5,
label=2,
attributes={
'truncated': True,
VOC.VocAction(1).name: True,
VOC.VocAction(2).name: True,
}),
Bbox(5,
4,
3,
2,
label=self._label('person'),
attributes={
'truncated': True,
VOC.VocAction(1).name: True,
VOC.VocAction(2).name: True,
}),
]),
])
def __iter__(self):
return iter([
DatasetItem(id=1, subset='train',
image=np.ones((16, 16, 3)),
annotations=[
Bbox(0, 4, 4, 8, label=0, group=2, id=0,
attributes={
'occluded': False, 'username': '',
}
),
Polygon([0, 4, 4, 4, 5, 6], label=1, id=1,
attributes={
'occluded': True, 'username': '',
'a1': 'qwe',
'a2': True,
'a3': 123,
}
),
Mask(np.array([[0, 1], [1, 0], [1, 1]]), group=2,
id=2, attributes={
'occluded': False, 'username': '******'
}
),
Bbox(1, 2, 3, 4, group=1, id=3, attributes={
'occluded': False, 'username': '',
}),
Mask(np.array([[0, 0], [0, 0], [1, 1]]), group=1,
id=4, attributes={
'occluded': True, 'username': ''
}
),
]
),
])
def test_can_save_dataset_with_no_subsets(self):
test_dataset = Dataset.from_iterable([
DatasetItem(id=1,
image=np.ones((16, 16, 3)),
annotations=[
Bbox(2, 1, 4, 4, label=2),
Bbox(4, 2, 8, 4, label=3),
],
attributes={'source_id': ''}
),
DatasetItem(id=2,
image=np.ones((8, 8, 3)) * 2,
annotations=[
Bbox(4, 4, 4, 4, label=3),
],
attributes={'source_id': ''}
),
DatasetItem(id=3,
image=np.ones((8, 4, 3)) * 3,
attributes={'source_id': ''}
),
], categories={
AnnotationType.label: LabelCategories.from_iterable(
'label_' + str(label) for label in range(10)),
})
with TestDir() as test_dir:
self._test_save_and_load(
test_dataset,
partial(TfDetectionApiConverter.convert, save_images=True),
test_dir)
def test_can_import(self):
target_dataset = Dataset.from_iterable([
DatasetItem(id=1, subset='train',
image=np.ones((16, 16, 3)),
annotations=[
Bbox(0, 4, 4, 8, label=2),
Bbox(0, 4, 4, 4, label=3),
Bbox(2, 4, 4, 4),
],
attributes={'source_id': '1'}
),
DatasetItem(id=2, subset='val',
image=np.ones((8, 8, 3)),
annotations=[
Bbox(1, 2, 4, 2, label=3),
],
attributes={'source_id': '2'}
),
DatasetItem(id=3, subset='test',
image=np.ones((5, 4, 3)) * 3,
attributes={'source_id': '3'}
),
], categories={
AnnotationType.label: LabelCategories.from_iterable(
'label_' + str(label) for label in range(10)),
})
dataset = Dataset.import_from(DUMMY_DATASET_DIR, 'tf_detection_api')
compare_datasets(self, target_dataset, dataset)
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 test_shapes_to_boxes(self):
source_dataset = Dataset.from_iterable([
DatasetItem(id=1, image=np.zeros((5, 5, 3)),
annotations=[
Mask(np.array([
[0, 0, 1, 1, 1],
[0, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 0],
[1, 1, 1, 0, 0]],
), id=1),
Polygon([1, 1, 4, 1, 4, 4, 1, 4], id=2),
PolyLine([1, 1, 2, 1, 2, 2, 1, 2], id=3),
Points([2, 2, 4, 2, 4, 4, 2, 4], id=4),
]
),
])
target_dataset = Dataset.from_iterable([
DatasetItem(id=1, image=np.zeros((5, 5, 3)),
annotations=[
Bbox(0, 0, 4, 4, id=1),
Bbox(1, 1, 3, 3, id=2),
Bbox(1, 1, 1, 1, id=3),
Bbox(2, 2, 2, 2, id=4),
]
),
])
actual = transforms.ShapesToBoxes(source_dataset)
compare_datasets(self, target_dataset, actual)
def test_can_load_dataset_with_exact_image_info(self):
source_dataset = Dataset.from_iterable(
[
DatasetItem(id=1,
subset='train',
image=Image(path='1.jpg', size=(10, 15)),
annotations=[
Bbox(0, 2, 4, 2, label=2),
Bbox(3, 3, 2, 3, label=4),
]),
],
categories={
AnnotationType.label:
LabelCategories.from_iterable('label_' + str(i)
for i in range(10)),
})
with TestDir() as test_dir:
YoloConverter.convert(source_dataset, test_dir)
parsed_dataset = Dataset.import_from(test_dir,
'yolo',
image_info={'1': (10, 15)})
compare_datasets(self, source_dataset, parsed_dataset)
def test_annotation_comparison(self):
a = Dataset.from_iterable([
DatasetItem(id=1, annotations=[
Caption('hello'), # unmatched
Caption('world', group=5),
Label(2, attributes={ 'x': 1, 'y': '2', }),
Bbox(1, 2, 3, 4, label=4, z_order=1, attributes={
'score': 1.0,
}),
Bbox(5, 6, 7, 8, group=5),
Points([1, 2, 2, 0, 1, 1], label=0, z_order=4),
Mask(label=3, z_order=2, image=np.ones((2, 3))),
]),
], categories=['a', 'b', 'c', 'd'])
b = Dataset.from_iterable([
DatasetItem(id=1, annotations=[
Caption('world', group=5),
Label(2, attributes={ 'x': 1, 'y': '2', }),
Bbox(1, 2, 3, 4, label=4, z_order=1, attributes={
'score': 1.0,
}),
Bbox(5, 6, 7, 8, group=5),
Bbox(5, 6, 7, 8, group=5), # unmatched
Points([1, 2, 2, 0, 1, 1], label=0, z_order=4),
Mask(label=3, z_order=2, image=np.ones((2, 3))),
]),
], categories=['a', 'b', 'c', 'd'])
comp = ExactComparator()
matched, unmatched, _, _, errors = comp.compare_datasets(a, b)
self.assertEqual(6, len(matched), matched)
self.assertEqual(2, len(unmatched), unmatched)
self.assertEqual(0, len(errors), errors)
def test_can_save_dataset_with_image_info(self):
source_dataset = Dataset.from_iterable(
[
DatasetItem(id=1,
subset='train',
image=Image(path='1.jpg', size=(10, 15)),
annotations=[
Bbox(0, 2, 4, 2, label=2),
Bbox(3, 3, 2, 3, label=4),
]),
],
categories={
AnnotationType.label:
LabelCategories.from_iterable('label_' + str(i)
for i in range(10)),
})
with TestDir() as test_dir:
YoloConverter.convert(source_dataset, test_dir)
save_image(osp.join(test_dir, 'obj_train_data', '1.jpg'),
np.ones((10, 15, 3))) # put the image for dataset
parsed_dataset = Dataset.import_from(test_dir, 'yolo')
compare_datasets(self, source_dataset, parsed_dataset)
def test_can_import(self):
expected_dataset = Dataset.from_iterable([
DatasetItem(id='0001_01', subset='train',
image=np.ones((10, 15, 3)),
annotations=[
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], label=0, group=1),
]
),
DatasetItem(id='0002_01', subset='train',
image=np.ones((10, 15, 3)),
annotations=[
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], label=1, group=1)
]
),
], categories={
AnnotationType.label: LabelCategories.from_iterable(
[('n000001', 'car'), ('n000002', 'person')]),
})
dataset = Dataset.import_from(DUMMY_DATASET_DIR, 'vgg_face2')
compare_datasets(self, expected_dataset, dataset)
def test_boxes_to_masks(self):
source_dataset = Dataset.from_iterable([
DatasetItem(id=1,
image=np.zeros((5, 5, 3)),
annotations=[
Bbox(0, 0, 3, 3, z_order=1),
Bbox(0, 0, 3, 1, z_order=2),
Bbox(0, 2, 3, 1, z_order=3),
]),
])
target_dataset = Dataset.from_iterable([
DatasetItem(
id=1,
image=np.zeros((5, 5, 3)),
annotations=[
Mask(np.array(
[[1, 1, 1, 0, 0], [1, 1, 1, 0, 0], [1, 1, 1, 0, 0],
[0, 0, 0, 0, 0], [0, 0, 0, 0, 0]], ),
z_order=1),
Mask(np.array(
[[1, 1, 1, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0],
[0, 0, 0, 0, 0], [0, 0, 0, 0, 0]], ),
z_order=2),
Mask(np.array(
[[0, 0, 0, 0, 0], [0, 0, 0, 0, 0], [1, 1, 1, 0, 0],
[0, 0, 0, 0, 0], [0, 0, 0, 0, 0]], ),
z_order=3),
]),
])
actual = transforms.BoxesToMasks(source_dataset)
compare_datasets(self, target_dataset, actual)
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 DISABLED_test_roi_nms():
ROI = namedtuple('ROI', ['conf', 'x', 'y', 'w', 'h', 'label'])
class_count = 3
noisy_count = 3
rois = [
ROI(0.3, 10, 40, 30, 10, 0),
ROI(0.5, 70, 90, 7, 10, 0),
ROI(0.7, 5, 20, 40, 60, 2),
ROI(0.9, 30, 20, 10, 40, 1),
]
pixel_jitter = 10
detections = []
for i, roi in enumerate(rois):
detections.append(
Bbox(roi.x,
roi.y,
roi.w,
roi.h,
label=roi.label,
attributes={'score': roi.conf}))
for j in range(noisy_count):
cls_conf = roi.conf * j / noisy_count
cls = (i + j) % class_count
box = [roi.x, roi.y, roi.w, roi.h]
offset = (np.random.rand(4) - 0.5) * pixel_jitter
detections.append(
Bbox(*(box + offset),
label=cls,
attributes={'score': cls_conf}))
import cv2
image = np.zeros((100, 100, 3))
for i, det in enumerate(detections):
roi = ROI(det.attributes['score'], *det.get_bbox(), det.label)
p1 = (int(roi.x), int(roi.y))
p2 = (int(roi.x + roi.w), int(roi.y + roi.h))
c = (0, 1 * (i % (1 + noisy_count) == 0), 1)
cv2.rectangle(image, p1, p2, c)
cv2.putText(image, 'd%s-%s-%.2f' % (i, roi.label, roi.conf), p1,
cv2.FONT_HERSHEY_SIMPLEX, 0.25, c)
cv2.imshow('nms_image', image)
cv2.waitKey(0)
nms_boxes = RISE.nms(detections, iou_thresh=0.25)
print(len(detections), len(nms_boxes))
for i, det in enumerate(nms_boxes):
roi = ROI(det.attributes['score'], *det.get_bbox(), det.label)
p1 = (int(roi.x), int(roi.y))
p2 = (int(roi.x + roi.w), int(roi.y + roi.h))
c = (0, 1, 0)
cv2.rectangle(image, p1, p2, c)
cv2.putText(image, 'p%s-%s-%.2f' % (i, roi.label, roi.conf), p1,
cv2.FONT_HERSHEY_SIMPLEX, 0.25, c)
cv2.imshow('nms_image', image)
cv2.waitKey(0)
def __iter__(self):
yield DatasetItem(id=1, annotations=[
Bbox(2, 3, 4, 5, label=0, id=1),
Bbox(1, 2, 3, 4, label=1, id=2, group=2,
attributes={'act1': True}),
Bbox(2, 3, 4, 5, label=2, id=3, group=2),
Bbox(2, 3, 4, 6, label=3, id=4, group=2),
])
def __iter__(self):
yield DatasetItem(id=1, annotations=[
Bbox(2, 3, 4, 5, label=self._label('foreign_label'), id=1),
Bbox(1, 2, 3, 4, label=self._label('label'), id=2, group=2,
attributes={'act1': True}),
Bbox(2, 3, 4, 5, label=self._label('label_part1'), group=2),
Bbox(2, 3, 4, 6, label=self._label('label_part2'), group=2),
])
def __iter__(self):
return iter([
DatasetItem(id=100,
subset='train',
image=np.ones((10, 6, 3)),
annotations=[
Caption('hello', id=1),
Caption('world', id=2, group=5),
Label(2, id=3, attributes={
'x': 1,
'y': '2',
}),
Bbox(1,
2,
3,
4,
label=4,
id=4,
z_order=1,
attributes={
'score': 1.0,
}),
Bbox(5, 6, 7, 8, id=5, group=5),
Points([1, 2, 2, 0, 1, 1],
label=0,
id=5,
z_order=4),
Mask(label=3,
id=5,
z_order=2,
image=np.ones((2, 3))),
]),
DatasetItem(id=21,
subset='train',
annotations=[
Caption('test'),
Label(2),
Bbox(1, 2, 3, 4, 5, id=42, group=42)
]),
DatasetItem(id=2,
subset='val',
annotations=[
PolyLine([1, 2, 3, 4, 5, 6, 7, 8],
id=11,
z_order=1),
Polygon([1, 2, 3, 4, 5, 6, 7, 8],
id=12,
z_order=4),
]),
DatasetItem(id=42,
subset='test',
attributes={
'a1': 5,
'a2': '42'
}),
DatasetItem(id=42),
DatasetItem(id=43, image=Image(path='1/b/c.qq', size=(2, 4))),
])
def __iter__(self):
return iter([
DatasetItem(id=1, subset='train',
image=Image(path='1.jpg', size=(10, 15)),
annotations=[
Bbox(0, 2, 4, 2, label=2),
Bbox(3, 3, 2, 3, label=4),
]),
])
def __iter__(self):
return iter([
DatasetItem(id=1,
subset='train',
image=np.ones((10, 15, 3)),
annotations=[
Bbox(0, 2, 4, 2, label=2),
Bbox(3, 3, 2, 3, label=4),
]),
])
def __iter__(self):
return iter([
DatasetItem(id=1,
image=np.zeros((5, 5, 3)),
annotations=[
Bbox(0, 0, 3, 3, z_order=1),
Bbox(0, 0, 3, 1, z_order=2),
Bbox(0, 2, 3, 1, z_order=3),
]),
])
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, label=2),
Bbox(0, 1, 2, 3, label=4),
]),
DatasetItem(id=2, subset='train', image=np.ones((10, 10, 3)),
annotations=[
Bbox(0, 2, 4, 2, label=2),
Bbox(3, 3, 2, 3, label=4),
Bbox(2, 1, 2, 3, label=4),
]),
DatasetItem(id=3, subset='valid', image=np.ones((8, 8, 3)),
annotations=[
Bbox(0, 1, 5, 2, label=2),
Bbox(0, 2, 3, 2, label=5),
Bbox(0, 2, 4, 2, label=6),
Bbox(0, 7, 3, 2, label=7),
]),
], categories={
AnnotationType.label: LabelCategories.from_iterable(
'label_' + str(i) for i in range(10)),
})
with TestDir() as test_dir:
YoloConverter.convert(source_dataset, test_dir, save_images=True)
parsed_dataset = YoloImporter()(test_dir).make_dataset()
compare_datasets(self, source_dataset, parsed_dataset)
def __iter__(self):
return iter([
DatasetItem(id=1,
subset='train',
image=np.ones((8, 8, 3)),
annotations=[
Bbox(0, 2, 4, 2, label=2),
Bbox(0, 1, 2, 3, label=4),
]),
DatasetItem(id=2,
subset='train',
image=np.ones((10, 10, 3)),
annotations=[
Bbox(0, 2, 4, 2, label=2),
Bbox(3, 3, 2, 3, label=4),
Bbox(2, 1, 2, 3, label=4),
]),
DatasetItem(id=3,
subset='valid',
image=np.ones((8, 8, 3)),
annotations=[
Bbox(0, 1, 5, 2, label=2),
Bbox(0, 2, 3, 2, label=5),
Bbox(0, 2, 4, 2, label=6),
Bbox(0, 7, 3, 2, label=7),
]),
])
def test_remap_labels(self):
src_dataset = Dataset.from_iterable(
[
DatasetItem(
id=1,
annotations=[
# Should be remapped
Label(1),
Bbox(1, 2, 3, 4, label=2),
Mask(image=np.array([1]), label=3),
# Should be kept
Polygon([1, 1, 2, 2, 3, 4], label=4),
PolyLine([1, 3, 4, 2, 5, 6])
])
],
categories={
AnnotationType.label:
LabelCategories.from_iterable('label%s' % i for i in range(5)),
AnnotationType.mask:
MaskCategories(colormap=mask_tools.generate_colormap(5)),
})
dst_dataset = Dataset.from_iterable(
[
DatasetItem(id=1,
annotations=[
Label(1),
Bbox(1, 2, 3, 4, label=0),
Mask(image=np.array([1]), label=1),
Polygon([1, 1, 2, 2, 3, 4], label=2),
PolyLine([1, 3, 4, 2, 5, 6], label=None)
]),
],
categories={
AnnotationType.label:
LabelCategories.from_iterable(['label0', 'label9', 'label4']),
AnnotationType.mask:
MaskCategories(
colormap={
k: v
for k, v in mask_tools.generate_colormap(5).items()
if k in {0, 1, 3, 4}
})
})
actual = transforms.RemapLabels(src_dataset,
mapping={
'label1': 'label9',
'label2': 'label0',
'label3': 'label9',
},
default='keep')
compare_datasets(self, dst_dataset, actual)
