def _create_matrices(self, shapes, random_state):
"""Create the transformation matrix
# Arguments
shapes [list of tuples]: list of image shapes
random_state [numpy Random state]: some random state
# Returns
[list of np array]: list of transformation matrices
[list of ints]: list of heights
[list of ints]: list of widths
"""
matrices = []
max_heights = []
max_widths = []
nb_images = len(shapes)
seeds = ia.copy_random_state(random_state).randint(
0, 10**6, (nb_images, ))
for _idx in range(nb_images):
height, width = shapes[_idx][:2]
pts1 = np.float32([[0, 0], [0, height - 1], [width - 1, 0],
[width - 1, height - 1]])
transition = self.jitter.draw_samples(
(4, 2), random_state=ia.new_random_state(seeds[_idx]))
transition[:, 0] = transition[:, 0] * np.min([height, width])
transition[:, 1] = transition[:, 1] * np.min([height, width])
transition = transition.astype(np.int32)
transition[:,
0] = transition[:, 0] + np.abs(np.min(transition[:, 0]))
transition[:,
1] = transition[:, 1] + np.abs(np.min(transition[:, 1]))
pts2 = np.float32(
[[transition[0, 0], transition[0, 1]],
[transition[1, 0], height - 1 + transition[1, 1]],
[width - 1 + transition[2, 0], transition[2, 1]],
[width - 1 + transition[3, 0],
height - 1 + transition[3, 1]]])
height = np.max(pts2[:, 1])
width = np.max(pts2[:, 0])
matrices.append(cv2.getPerspectiveTransform(pts1, pts2))
max_heights.append(height)
max_widths.append(width)
return matrices, max_heights, max_widths
def augment_object_data(object_data,
random_state=None,
fit_output=True,
aug_color=True,
aug_geo=True,
augmentations=None,
random_order=False,
scale=(0.5, 1.0)):
try:
iaa.Affine(fit_output=True)
except TypeError:
warnings.warn(
'Your imgaug does not support fit_output kwarg for'
'imgaug.augmenters.Affine. Please install via'
'\n\n\tpip install -e git+https://github.com/wkentaro/imgaug@affine_resize\n\n' # NOQA
'to enable it.')
fit_output = False
if random_state is None:
random_state = np.random.RandomState()
if augmentations is None:
st = lambda x: iaa.Sometimes(0.3, x) # NOQA
kwargs_affine = dict(
order=1, # order=0 for mask
cval=0,
scale=scale,
translate_px=(-16, 16),
rotate=(-180, 180),
shear=(-16, 16),
mode='constant',
)
if fit_output:
kwargs_affine['fit_output'] = fit_output
augmentations = [
st(
iaa.WithChannels([0, 1], iaa.Multiply([1, 1.5])
) if aug_color else iaa.Noop()),
st(
iaa.WithColorspace('HSV',
children=iaa.
WithChannels([1, 2], iaa.Multiply([0.5, 2]))
) if aug_color else iaa.Noop()),
st(
iaa.GaussianBlur(
sigma=[0.0, 1.0]) if aug_color else iaa.Noop()),
iaa.Sometimes(
0.8,
iaa.Affine(**kwargs_affine) if aug_geo else iaa.Noop()),
]
aug = iaa.Sequential(
augmentations,
random_order=random_order,
random_state=ia.copy_random_state(random_state),
)
def activator_imgs(images, augmenter, parents, default):
if isinstance(augmenter, iaa.Affine):
augmenter.order = Deterministic(1)
augmenter.cval = Deterministic(0)
return True
def activator_masks(images, augmenter, parents, default):
white_lists = (iaa.Affine, iaa.PerspectiveTransform, iaa.Sequential,
iaa.Sometimes)
if not isinstance(augmenter, white_lists):
return False
if isinstance(augmenter, iaa.Affine):
augmenter.order = Deterministic(0)
augmenter.cval = Deterministic(0)
return True
def activator_lbls(images, augmenter, parents, default):
white_lists = (iaa.Affine, iaa.PerspectiveTransform, iaa.Sequential,
iaa.Sometimes)
if not isinstance(augmenter, white_lists):
return False
if isinstance(augmenter, iaa.Affine):
augmenter.order = Deterministic(0)
augmenter.cval = Deterministic(-1)
return True
for objd in object_data:
aug = aug.to_deterministic()
objd['img'] = aug.augment_image(
objd['img'], hooks=ia.HooksImages(activator=activator_imgs))
if 'mask' in objd:
objd['mask'] = aug.augment_image(
objd['mask'], hooks=ia.HooksImages(activator=activator_masks))
if 'lbl' in objd:
objd['lbl'] = aug.augment_image(
objd['lbl'], hooks=ia.HooksImages(activator=activator_lbls))
if 'lbl_suc' in objd:
objd['lbl_suc'] = aug.augment_image(
objd['lbl_suc'],
hooks=ia.HooksImages(activator=activator_lbls))
if 'masks' in objd:
masks = []
for mask in objd['masks']:
mask = aug.augment_image(
mask,
hooks=ia.HooksImages(activator=activator_masks),
)
masks.append(mask)
masks = np.asarray(masks)
objd['masks'] = masks
del masks
if 'lbls' in objd:
lbls = []
for lbl in objd['lbls']:
lbl = aug.augment_image(
lbl,
hooks=ia.HooksImages(activator=activator_lbls),
)
lbls.append(lbl)
lbls = np.asarray(lbls)
objd['lbls'] = lbls
del lbls
yield objd