def scale(img, size):
"""Rescales the input image to the given "size".
If height of the image is larger than its width,
image will be resized to (size * height / width, size).
Similar resizing will be done otherwise.
Args:
img (numpy.ndarray): an image array
size (int): the length of the smaller edge.
Returns:
~numpy.ndarray: A scaled image
"""
_, H, W = img.shape
if (W <= H and W == size) or (H <= W and H == size):
return img
if W < H:
oH = int(size * H / W)
return resize(img, (oH, size))
else:
oW = int(size * W / H)
return resize(img, (size, oW))
def stack_imgs(fns, crop, resize=False, grey=False):
imgs_in = []
for fn in fns:
fn1, ext = os.path.splitext(fn)
# image can be given as csv or jpg/png... etc
if ext == ".csv":
img_in = np.loadtxt(fn, delimiter=",")[np.newaxis, ]
elif ext == ".npy":
img_in = (np.load(fn)[np.newaxis, ]).astype(np.float32)
img_in = (np.sqrt(np.clip(img_in, 0,
100))) / 10.0 ## nasty preprocess
# img_in = (img_in - np.mean(img_in))/2*np.std(img_in) # standardize
else:
img_in = read_image(fn, color=not grey) / 127.5 - 1.0
# resize if the image is too small
if resize:
if img_in.shape[1] < crop[0] or img_in.shape[2] < crop[1]:
if crop[0] / img_in.shape[1] < crop[1] / img_in.shape[2]:
img_in = resize(img_in, (int(
crop[1] / img_in.shape[2] * img_in.shape[1]), crop[1]))
else:
img_in = resize(
img_in,
(crop[0],
int(crop[0] / img_in.shape[1] * img_in.shape[2])))
imgs_in.append(img_in)
# an input/output image can consist of multiple images; they are stacked as channels
# print(imgs_in.shape)
return (np.concatenate(imgs_in, axis=0))
def _get_proba(self, img, scale, flip):
if flip:
img = img[:, :, ::-1]
_, H, W = img.shape
if scale == 1.0:
h, w = H, W
else:
h, w = int(H * scale), int(W * scale)
img = resize(img, (h, w))
img = self.prepare(img)
x = chainer.Variable(self.xp.asarray(img[np.newaxis]))
x = self.__call__(x)
x = F.softmax(x, axis=1)
score = F.resize_images(x, img.shape[1:])[0, :, :h, :w].array
score = chainer.backends.cuda.to_cpu(score)
if scale != 1.0:
score = resize(score, (H, W))
if flip:
score = score[:, :, ::-1]
return score
def __call__(self, in_data):
assert len(in_data) == 6
img, bbox, label, mask, lbl_vis, lbl_occ = in_data
# H, W, C -> C, H, W
img = img.transpose(2, 0, 1)
lbl_occ = lbl_occ.transpose(2, 0, 1)
if not self.train:
return img, bbox, label, mask, lbl_vis, lbl_occ
imgs, sizes, scales = self.mask_rcnn.prepare([img])
img = imgs[0]
H, W = sizes[0]
scale = scales[0]
# _, o_H, o_W = img.shape
o_H, o_W = int(round(scale * H)), int(round(scale * W))
if len(bbox) > 0:
bbox = transforms.resize_bbox(bbox, (H, W), (o_H, o_W))
if len(mask) > 0:
mask = transforms.resize(mask, size=(o_H, o_W), interpolation=0)
mask = mask.transpose(1, 2, 0)
mask = pad_multiple_of(mask, mode='constant', constant_values=-1)
mask = mask.transpose(2, 0, 1)
assert mask.shape[1:] == img.shape[1:]
lbl_vis = transforms.resize(lbl_vis[None],
size=(o_H, o_W),
interpolation=0)[0]
lbl_occ = transforms.resize(lbl_occ, size=(o_H, o_W), interpolation=0)
lbl_vis = pad_multiple_of(lbl_vis, mode='constant', constant_values=-1)
lbl_occ = lbl_occ.transpose(1, 2, 0)
lbl_occ = pad_multiple_of(lbl_occ, mode='constant', constant_values=-1)
lbl_occ = lbl_occ.transpose(2, 0, 1)
assert lbl_vis.shape == img.shape[1:]
assert lbl_occ.shape[1:] == img.shape[1:]
# # horizontally flip
# img, params = transforms.random_flip(
# img, x_random=True, return_param=True)
# bbox = transforms.flip_bbox(
# bbox, (o_H, o_W), x_flip=params['x_flip'])
# if mask.ndim == 2:
# mask = transforms.flip(
# mask[None, :, :], x_flip=params['x_flip'])[0]
# else:
# mask = transforms.flip(mask, x_flip=params['x_flip'])
# lbl_vis = transforms.flip(lbl_vis[None], x_flip=params['x_flip'])[0]
# lbl_occ = transforms.flip(lbl_occ, x_flip=params['x_flip'])
keep = (mask == 1).sum(axis=(1, 2)) > 0
bbox = bbox[keep]
label = label[keep]
mask = mask[keep]
return img, bbox, label, mask, scale, lbl_vis, lbl_occ
def __call__(self, in_data):
img, label = in_data
_, height, width = img.shape
scale = np.random.uniform(self.scale_range[0], self.scale_range[1])
# Scale
scaled_height = int(scale * height)
scaled_width = int(scale * width)
img = transforms.resize(img, (scaled_height, scaled_width),
PIL.Image.BICUBIC)
label = transforms.resize(label[None], (scaled_height, scaled_width),
PIL.Image.NEAREST)[0]
# Crop
if (scaled_height < self.crop_size[0]) or (scaled_width <
self.crop_size[1]):
shorter_side = min(img.shape[1:])
img, param = transforms.random_crop(img,
(shorter_side, shorter_side),
True)
else:
img, param = transforms.random_crop(img, self.crop_size, True)
label = label[param['y_slice'], param['x_slice']]
# Rotate
angle = np.random.uniform(-10, 10)
img = transforms.rotate(img, angle, expand=False)
label = transforms.rotate(label[None],
angle,
expand=False,
interpolation=PIL.Image.NEAREST,
fill=-1)[0]
# Resize
if ((img.shape[1] < self.crop_size[0])
or (img.shape[2] < self.crop_size[1])):
img = transforms.resize(img, self.crop_size, PIL.Image.BICUBIC)
if ((label.shape[0] < self.crop_size[0])
or (label.shape[1] < self.crop_size[1])):
label = transforms.resize(label[None].astype(np.float32),
self.crop_size, PIL.Image.NEAREST)
label = label.astype(np.int32)[0]
# heightorizontal flip
if self.horizontal_flip and np.random.rand() > 0.5:
img = transforms.flip(img, x_flip=True)
label = transforms.flip(label[None], x_flip=True)[0]
# Mean subtraction
img = img - self.mean
return img, label
def stack_imgs(self, fns, resize=False, onehot=False, clip=(None, None)):
imgs_in = []
for fn in fns:
fn1, ext = os.path.splitext(fn)
# image can be given as csv or jpg/png... etc
if ext == ".csv":
img_in = np.loadtxt(fn, delimiter=",")
elif ext == ".txt":
img_in = np.loadtxt(fn)
elif ext == ".npy":
img_in = np.load(fn)
# img_in = (np.sqrt(np.clip(img_in,0,100)))/10.0 ## nasty preprocess
# img_in = (img_in - np.mean(img_in))/2*np.std(img_in) # standardize
elif ext == ".dcm":
ref_dicom_in = dicom.read_file(fn, force=True)
ref_dicom_in.file_meta.TransferSyntaxUID = dicom.uid.ImplicitVRLittleEndian
img_in = ref_dicom_in.pixel_array + ref_dicom_in.RescaleIntercept
else: ## image file
img_in = read_image(fn, color=not self.grey)
# make the image shape to [C,H,W]
if len(img_in.shape) == 2:
img_in = img_in[np.newaxis, ]
# resize if the image is too small
if resize:
if img_in.shape[1] < self.crop[0] or img_in.shape[
2] < self.crop[1]:
if self.crop[0] / img_in.shape[1] < self.crop[
1] / img_in.shape[2]:
img_in = resize(img_in,
(int(self.crop[1] / img_in.shape[2] *
img_in.shape[1]), self.crop[1]))
else:
img_in = resize(img_in,
(self.crop[0],
int(self.crop[0] / img_in.shape[1] *
img_in.shape[2])))
imgs_in.append(img_in)
imgs_in = np.concatenate(imgs_in, axis=0)
# print(imgs_in.shape)
if onehot > 0:
return (np.eye(self.class_num)[imgs_in[0].astype(
np.uint64)].astype(np.float32).transpose((2, 0, 1)))
else:
## clip and normalise to [-1,1]
if clip[0] is not None:
imgs_in = np.clip(imgs_in, clip[0], clip[1])
imgs_in = 2 * (imgs_in - clip[0]) / (clip[1] - clip[0]) - 1.0
return (imgs_in.astype(np.float32))
def transform_img(inputs,
mean,
std,
pca_sigma=0,
random_angle=0,
x_random_flip=False,
y_random_flip=False,
expand_ratio=1.,
random_crop_size=(224, 224),
random_erase=False,
output_size=(224, 224),
train=False):
x, lab = inputs
x = x.copy()
# Color augmentation
if train and pca_sigma != 0:
x = transforms.pca_lighting(x, pca_sigma)
x -= mean[:, None, None]
x /= std[:, None, None]
x = x[::-1]
if train:
# Random rotate
if random_angle != 0:
angle = np.random.uniform(-random_angle, random_angle)
x = cv_rotate(x, angle)
# Random flip
if x_random_flip or y_random_flip:
x = transforms.random_flip(x,
x_random=x_random_flip,
y_random=y_random_flip)
# Random expand
if expand_ratio > 1:
x = transforms.random_expand(x, max_ratio=expand_ratio)
if all(random_crop_size) > 0:
x = transforms.random_crop(x, random_crop_size)
else:
if random_erase:
x = random_erasing(x)
if all(random_crop_size) > 0:
x = transforms.resize(x, random_crop_size)
else:
x = transforms.resize(x, output_size)
return x, lab
def get_example(self, i):
if self.imgtype == "npy":
img = np.load(self.get_img_path(i))
img = 2 * (np.clip(img, self.base, self.base + self.range) -
self.base) / self.range - 1.0
if len(img.shape) == 2:
img = img[np.newaxis, ]
else:
ref_dicom = dicom.read_file(self.get_img_path(i), force=True)
# print(ref_dicom)
# ref_dicom.file_meta.TransferSyntaxUID = dicom.uid.ImplicitVRLittleEndian
img = ref_dicom.pixel_array + ref_dicom.RescaleIntercept
img = self.img2var(img)
img = img[np.newaxis, :, :]
if self.scale_to > 0:
img = resize(img, (self.scale_to, self.scale_to))
H, W = self.crop
# print(img.shape)
if img.shape[1] < H + 2 * self.random or img.shape[
2] < W + 2 * self.random:
p = max(H + 2 * self.random - img.shape[1],
W + 2 * self.random - img.shape[2])
img = np.pad(img, ((0, 0), (p, p), (p, p)), 'edge')
if H + self.random < img.shape[1] and W + self.random < img.shape[2]:
img = center_crop(img, (H + self.random, W + self.random))
img = random_crop(img, self.crop)
return img
def predict(self, imgs):
"""Conduct semantic segmentations from images.
Args:
imgs (iterable of numpy.ndarray): Arrays holding images.
All images are in CHW and RGB format
and the range of their values are :math:`[0, 255]`.
Returns:
list of numpy.ndarray:
List of integer labels predicted from each image in the input \
list.
"""
labels = []
for img in imgs:
C, H, W = img.shape
with chainer.using_config('train', False), \
chainer.function.no_backprop_mode():
x = chainer.Variable(self.xp.asarray(img[np.newaxis]))
score = self.__call__(x)[0].data
score = chainer.cuda.to_cpu(score)
if score.shape != (C, H, W):
dtype = score.dtype
score = resize(score, (H, W)).astype(dtype)
label = np.argmax(score, axis=0).astype(np.int32)
labels.append(label)
return labels
def _transform2(data, mean, train=True, mean_flag=False):
img, label = data
img = img.copy()
size316 = (316, 316)
size = (224, 224)
img_o = transforms.scale(img, 316)
img_o = transforms.center_crop(img_o, size316)
# 学習のときだけ実行
if train:
img_o = transforms.random_flip(img_o, y_random=True)
img_o = transforms.random_rotate(img_o)
# img = random_erase(img)
img_o = transforms.resize(img_o, size)
# 画像から平均を引く
if mean_flag:
img_o -= mean
img_o *= (1.0 / 255.0)
r = random.randint(316, 1500)
img_st = transforms.scale(img, r)
img_st = transforms.center_crop(img_st, (224, 224))
# 画像から平均を引く
if mean_flag:
img_st -= mean
img_st *= (1.0 / 255.0)
return img_o, label, img_st
def __call__(self, in_data):
img, label = in_data
img = random_sized_crop(img)
img = resize(img, (224, 224))
img = random_flip(img, x_random=True)
img -= self.mean
return img, label
def __call__(self, in_data):
if len(in_data) == 4:
img, mask, label, bbox = in_data
else:
img, bbox, label = in_data
# Flipping
img, params = transforms.random_flip(img,
x_random=True,
return_param=True)
x_flip = params['x_flip']
bbox = transforms.flip_bbox(bbox, img.shape[1:], x_flip=x_flip)
# Scaling and mean subtraction
img, scale = scale_img(img, self.min_size, self.max_size)
img -= self.mean
bbox = bbox * scale
if len(in_data) == 4:
mask = transforms.flip(mask, x_flip=x_flip)
mask = transforms.resize(mask.astype(np.float32),
img.shape[1:],
interpolation=PIL.Image.NEAREST).astype(
np.bool)
return img, bbox, label, mask
else:
return img, bbox, label
def transform(data, mean, train=True):
img, lable = data
img = img.copy()
img -= mean
size = (224, 224)
if train:
h, w = img.shape[1:]
angles = [i for i in range(0, 360, 10)]
angle = np.random.choice(angles)
img = rotate(img, angle)
rad = angle * np.pi / 180
new_length = int(h / (np.abs(np.cos(rad)) + np.abs(np.sin(rad))))
img = transforms.center_crop(img, (new_length, new_length))
# img = transforms.random_rotate(img, return_param=False)
img = transforms.random_flip(img, x_random=True)
img = transforms.resize(img, size, interpolation=2)
img *= (1.0 / 255.0)
return img, lable
def _image_process(self, img):
img = img / 255
img = resize(img, (224, 224))
if np.random.rand() >= 0.5:
img = img[:, :, ::-1]
return img
def __call__(self, in_data):
# There are five data augmentation steps
# 3. Random cropping
# 4. Resizing with random interpolation
# 5. Random horizontal flipping
img, bbox, label = in_data
# 3. Random cropping
if self.random_crop and np.random.rand() > 0.5:
next_img, param = random_crop_with_bbox_constraints(
img,
bbox,
min_scale=min(self.crop_rate),
max_scale=max(self.crop_rate),
return_param=True)
next_bbox, param = transforms.crop_bbox(bbox,
y_slice=param['y_slice'],
x_slice=param['x_slice'],
allow_outside_center=False,
return_param=True)
if (len(label[param['index']]) != 0):
label = label[param['index']]
img, bbox = next_img, next_bbox
# 4. Resizing with random interpolatation
_, H, W = img.shape
img = transforms.resize(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. Random horizontal flipping
if self.flip:
img, params = transforms.random_flip(img,
x_random=True,
return_param=True)
bbox = transforms.flip_bbox(bbox, (self.size, self.size),
x_flip=params['x_flip'])
img -= self.mean
img /= self.std
_, height, width = img.shape
ymin = bbox[:, 0]
xmin = bbox[:, 1]
ymax = bbox[:, 2]
xmax = bbox[:, 3]
one_hot_label = np.eye(self.n_class)[label]
xs = (xmin + (xmax - xmin) // 2) / width
ws = (xmax - xmin) / width
ys = (ymin + (ymax - ymin) // 2) / height
hs = (ymax - ymin) / height
t = [{
'label': l,
'x': x,
'w': w,
'y': y,
'h': h,
'one_hot_label': hot
} for l, x, w, y, h, hot in zip(label, xs, ws, ys, hs, one_hot_label)]
return img, t
def _prepare(self, img):
"""Prepare an image for feeding it to a model.
This is a standard preprocessing scheme used by feature extraction
models.
First, the image is scaled or resized according to :math:`scale_size`.
Note that this step is optional.
Next, the image is cropped to :math:`crop_size`.
Last, the image is mean subtracted by an array :obj:`mean`.
Args:
img (~numpy.ndarray): An image. This is in CHW format.
The range of its value is :math:`[0, 255]`.
Returns:
~numpy.ndarray:
A preprocessed image. This is 4D array whose batch size is
the number of crops.
"""
if self.scale_size is not None:
if isinstance(self.scale_size, int):
img = scale(img, size=self.scale_size)
else:
img = resize(img, size=self.scale_size)
if self.crop == '10':
imgs = ten_crop(img, self.crop_size)
elif self.crop == 'center':
imgs = center_crop(img, self.crop_size)[np.newaxis]
imgs -= self.mean[np.newaxis]
return imgs
def predict(self, imgs):
"""Conduct semantic segmentations from images.
Args:
imgs (iterable of numpy.ndarray): Arrays holding images.
All images are in CHW and RGB format
and the range of their values are :math:`[0, 255]`.
Returns:
list of numpy.ndarray:
List of integer labels predicted from each image in the input \
list.
"""
labels = []
for img in imgs:
C, H, W = img.shape
with chainer.using_config('train', False), \
chainer.function.no_backprop_mode():
x = chainer.Variable(self.xp.asarray(img[np.newaxis]))
score = self.__call__(x)[0].data
score = chainer.cuda.to_cpu(score)
if score.shape != (C, H, W):
dtype = score.dtype
score = resize(score, (H, W)).astype(dtype)
label = np.argmax(score, axis=0).astype(np.int32)
labels.append(label)
return labels
def __call__(self, in_data):
img, label = in_data
img = transforms.random_sized_crop(img)
img = transforms.resize(img, (224, 224))
img = transforms.random_flip(img, x_random=True)
img -= self.mean
return img.astype(chainer.get_dtype()), label
def _prepare(self, img):
"""Prepare an image for feeding it to a model.
This is a standard preprocessing scheme used by feature extraction
models.
First, the image is scaled or resized according to :math:`scale_size`.
Note that this step is optional.
Next, the image is cropped to :math:`crop_size`.
Last, the image is mean subtracted by an array :obj:`mean`.
Args:
img (~numpy.ndarray): An image. This is in CHW format.
The range of its value is :math:`[0, 255]`.
Returns:
~numpy.ndarray:
A preprocessed image. This is 4D array whose batch size is
the number of crops.
"""
if self.scale_size is not None:
if isinstance(self.scale_size, int):
img = scale(img, size=self.scale_size)
else:
img = resize(img, size=self.scale_size)
if self.crop == '10':
imgs = ten_crop(img, self.crop_size)
elif self.crop == 'center':
imgs = center_crop(img, self.crop_size)[np.newaxis]
imgs -= self.mean[np.newaxis]
return imgs
def _transform2(data, mean, train=True, mean_flag=False):
img, label = data
img = img.copy()
size316 = (316, 316)
size = (224, 224)
img = transforms.scale(img, 316)
img = transforms.center_crop(img, size316)
# 学習のときだけ実行
if train:
img = transforms.random_flip(img, y_random=True)
img = transforms.random_rotate(img)
# img = random_erase(img)
img = transforms.resize(img, size)
img = img.transpose(1, 2, 0)
# 画像から平均を引く
if mean_flag:
img -= mean
img *= (1.0 / 255.0)
img = img.transpose(2, 0, 1)
return img, label
def test_zero_length_img(self):
if self.backend == 'cv2' and not _cv2_available:
return
img = np.random.uniform(size=(0, 24, 32))
with chainer.using_config('cv_resize_backend', self.backend):
out = resize(img, size=(32, 64), interpolation=self.interpolation)
self.assertEqual(out.shape, (0, 32, 64))
def random_resize(img):
rv = random.random()
if rv < 0.5:
ratio = round(rv * 2, 1)
_, H, W = img.shape
img = transforms.resize(img, (int(ratio * H), int(ratio * W)))
return img
def gen_morphed_images(z, base_class, palette, masks, interpolation=8):
z = xp.broadcast_to(z, (interpolation, 128))
sizes = [4, 8, 8, 16, 16, 32, 32, 64, 64, 128, 128, 256]
ws = []
for i_size, size in enumerate(sizes):
w = xp.zeros((interpolation, size, size, gen.n_classes),
dtype=xp.float32)
w[:, :, :, base_class] = 1.0 # default class
for i_mask in range(len(palette)):
resized_mask = xp.array(resize(masks[i_mask],
(size, size)).reshape((size, size)),
dtype=xp.float32)
# resized_mask = xp.array(img_masks[i_mask].resize((size, size))).astype(xp.float32) / 255
for i in range(interpolation):
weight = i / (interpolation - 1.0)
# if i_size <= 0:
# weight = 0
w[i, :, :, base_class] -= resized_mask * weight
w[i, :, :, palette[i_mask]] = resized_mask * weight
ws.append(chainer.Variable(w))
with chainer.using_config('train', False), chainer.using_config(
'enable_backprop', False):
x = gen.spatial_interpolation(z, ws)
x = x.data
if args.gpu >= 0:
x = x.get()
x = np.asarray(np.clip(x * 127.5 + 127.5, 0.0, 255.0),
dtype=np.uint8).transpose((0, 2, 3, 1))
return x
def segm_to_mask(segm, bbox, size):
"""Recover mask from cropped and resized mask.
This function requires cv2.
Args:
segm (~numpy.ndarray): See below.
bbox (~numpy.ndarray): See below.
size (tuple): This is a tuple of length 2. Its elements are
ordered as (height, width).
Returns:
~numpy.ndarray: See below.
.. csv-table::
:header: name, shape, dtype, format
:obj:`segm`, ":math:`(R, S, S)`", :obj:`float32`, --
:obj:`bbox`, ":math:`(R, 4)`", :obj:`float32`, \
":math:`(y_{min}, x_{min}, y_{max}, x_{max})`"
:obj:`mask` (output), ":math:`(R, H, W)`", :obj:`bool`, --
"""
pad = 1
H, W = size
_, segm_size, _ = segm.shape
mask = np.zeros((len(bbox), H, W), dtype=np.bool)
# As commented in mask_to_segm, cv2.resize needs adjust.
padded_segm_size = segm_size + pad * 2
expand_scale = padded_segm_size / segm_size
bbox = _expand_bbox(bbox, expand_scale)
canvas_mask = np.zeros((padded_segm_size, padded_segm_size),
dtype=np.float32)
bbox = _integerize_bbox(bbox)
for i, (bb, sgm) in enumerate(zip(bbox, segm)):
bb_height = bb[2] - bb[0]
bb_width = bb[3] - bb[1]
if bb_height == 0 or bb_width == 0:
continue
canvas_mask[pad:-pad, pad:-pad] = sgm
with chainer.using_config('cv_resize_backend', 'cv2'):
crop_mask = transforms.resize(canvas_mask[None],
(bb_height, bb_width))[0]
crop_mask = crop_mask > 0.5
y_min = max(bb[0], 0)
x_min = max(bb[1], 0)
y_max = max(min(bb[2], H), 0)
x_max = max(min(bb[3], W), 0)
y_offset = y_min - bb[0]
x_offset = x_min - bb[1]
mask[i, y_min:y_max,
x_min:x_max] = crop_mask[y_offset:y_offset + y_max - y_min,
x_offset:x_offset + x_max - x_min]
return mask
def scale_mask(mask, bbox, size):
"""Scale instance segmentation mask while keeping the aspect ratio.
This function exploits the sparsity of :obj:`mask` to speed up
resize operation.
The input image will be resized so that
the shorter edge will be scaled to length :obj:`size` after
resizing.
Args:
mask (array): An array whose shape is :math:`(R, H, W)`.
:math:`R` is the number of masks.
The dtype should be :obj:`numpy.bool`.
bbox (array): The bounding boxes around the masked region
of :obj:`mask`. This is expected to be the value
obtained by :obj:`bbox = chainercv.utils.mask_to_bbox(mask)`.
size (int): The length of the smaller edge.
Returns:
array:
An array whose shape is :math:`(R, H, W)`.
:math:`R` is the number of masks.
The dtype should be :obj:`numpy.bool`.
"""
xp = chainer.backends.cuda.get_array_module(mask)
mask = chainer.cuda.to_cpu(mask)
bbox = chainer.cuda.to_cpu(bbox)
R, H, W = mask.shape
if H < W:
out_size = (size, int(size * W / H))
scale = size / H
else:
out_size = (int(size * H / W), size)
scale = size / W
bbox[:, :2] = np.floor(bbox[:, :2])
bbox[:, 2:] = np.ceil(bbox[:, 2:])
bbox = bbox.astype(np.int32)
scaled_bbox = bbox * scale
scaled_bbox[:, :2] = np.floor(scaled_bbox[:, :2])
scaled_bbox[:, 2:] = np.ceil(scaled_bbox[:, 2:])
scaled_bbox = scaled_bbox.astype(np.int32)
out_mask = xp.zeros((R, ) + out_size, dtype=np.bool)
for i, (m, bb, scaled_bb) in enumerate(zip(mask, bbox, scaled_bbox)):
cropped_m = m[bb[0]:bb[2], bb[1]:bb[3]]
h = scaled_bb[2] - scaled_bb[0]
w = scaled_bb[3] - scaled_bb[1]
cropped_m = transforms.resize(cropped_m[None].astype(np.float32),
(h, w),
interpolation=PIL.Image.NEAREST)[0]
if xp != np:
cropped_m = xp.array(cropped_m)
out_mask[i, scaled_bb[0]:scaled_bb[2],
scaled_bb[1]:scaled_bb[3]] = cropped_m
return out_mask
def transform(inputs, mean, std, output_size=(224, 224)):
x, lab = inputs
x = x.copy()
x -= mean[:, None, None]
x /= std[:, None, None]
x = x[::-1]
x = transforms.resize(x, output_size)
return x, lab
def _transform(inputs, mean=None, img_size=(512, 1024), scale_label=1):
img, label = inputs
# Scaling
if img_size:
img_size = (img_size[0], img_size[1])
img = transforms.resize(img, img_size, Image.BICUBIC)
label = transforms.resize(label[None, ...], img_size, Image.NEAREST)[0]
# Mean subtraction
if mean is not None:
img -= mean[:, None, None]
if scale_label != 1:
scale_label = (int(label.shape[1]/scale_label),\
int(label.shape[0]/scale_label))
label = cv.resize(label, scale_label, interpolation=cv.INTER_NEAREST)
return img, label
def __call__(self, in_data):
if len(in_data) == 6:
img, bbox, label, mask, crowd, area = in_data
elif len(in_data) == 4:
img, bbox, label, mask = in_data
else:
raise ValueError
img = img.transpose(2, 0, 1) # H, W, C -> C, H, W
if not self.train:
if len(in_data) == 6:
return img, bbox, label, mask, crowd, area
elif len(in_data) == 4:
return img, bbox, label, mask
else:
raise ValueError
imgs, sizes, scales = self.mask_rcnn.prepare([img])
# print(type(imgs))
# print(type(sizes))
# print(type(scales))
img = imgs[0]
H, W = sizes[0]
scale = scales[0]
_, o_H, o_W = img.shape
if len(bbox) > 0:
bbox = transforms.resize_bbox(bbox, (H, W), (o_H, o_W))
if len(mask) > 0:
mask = transforms.resize(
mask, size=(o_H, o_W), interpolation=0)
# # horizontally flip
# img, params = transforms.random_flip(
# img, x_random=True, return_param=True)
# bbox = transforms.flip_bbox(
# bbox, (o_H, o_W), x_flip=params['x_flip'])
# if mask.ndim == 2:
# mask = transforms.flip(
# mask[None, :, :], x_flip=params['x_flip'])[0]
# else:
# mask = transforms.flip(mask, x_flip=params['x_flip'])
# horizontally and vartically flip
img, params = transforms.random_flip(
img, y_random=True, x_random=True, return_param=True)
bbox = transforms.flip_bbox(
bbox, (o_H, o_W), y_flip=params['y_flip'], x_flip=params['x_flip'])
if mask.ndim == 2:
mask = transforms.flip(
mask[None, :, :], y_flip=params['y_flip'], x_flip=params['x_flip'])[0]
else:
mask = transforms.flip(mask, y_flip=params['y_flip'], x_flip=params['x_flip'])
return img, bbox, label, mask, scale, sizes
def scale_img(img, min_size, max_size):
"""Process image."""
_, H, W = img.shape
scale = min_size / min(H, W)
if scale * max(H, W) > max_size:
scale = max_size / max(H, W)
H, W = int(H * scale), int(W * scale)
img = transforms.resize(img, (H, W))
return img, scale
def _scale_img(self, img):
"""Process image."""
_, H, W = img.shape
scale = self._min_size / min(H, W)
if scale * max(H, W) > self._max_size:
scale = self._max_size / max(H, W)
H, W = int(H * scale), int(W * scale)
img = transforms.resize(img, (H, W))
return img, scale
def test_transform(sample):
img = sample
if len(img.shape) == 2: # Grayscale
img = np.stack([img, img, img], 2)
img = np.transpose(img, (2, 0, 1))
img = transforms.resize(img, resize_size)
img = transforms.center_crop(img, patchsize)
img = img - mean
img = img.astype(dtype)
return img
def _prepare(self, img):
img = img.astype(np.float32)
img = transforms.resize(img, (self.insize, self.insize))
img -= self.mean
# NOTE: chainer.get_dtype will return float16 if the
# global_config.dtype is chainer.mixed16
img = img.astype(chainer.get_dtype())
return img
def __call__(self, imgs):
resized_imgs = []
for img in imgs:
_, H, W = img.shape
scale = self._min_size / min(H, W)
if scale * max(H, W) > self._max_size:
scale = self._max_size / max(H, W)
H, W = int(H * scale), int(W * scale)
img = transforms.resize(img, (H, W))
img -= self._mean
resized_imgs.append(img)
size = np.array([img.shape[1:] for img in resized_imgs]).max(axis=0)
size = (np.ceil(size / self._stride) * self._stride).astype(int)
x = np.zeros((len(imgs), 3, size[0], size[1]), dtype=np.float32)
for i, img in enumerate(resized_imgs):
_, H, W = img.shape
x[i, :, :H, :W] = img
return x
def test_resize_grayscale(self):
img = np.random.uniform(size=(1, 24, 32))
out = resize(img, size=(32, 64), interpolation=self.interpolation)
self.assertEqual(out.shape, (1, 32, 64))
def transform(in_data):
img, label = in_data
img = resize(img, (32, 32))
return img, label
def resize_contain(img, size, fill=0, interpolation=PIL.Image.BILINEAR,
return_param=False):
"""Resize the image to fit in the given area while keeping aspect ratio.
If both the height and the width in :obj:`size` are larger than
the height and the width of the :obj:`img`, the :obj:`img` is placed on
the center with an appropriate padding to match :obj:`size`.
Otherwise, the input image is scaled to fit in a canvas whose size
is :obj:`size` while preserving aspect ratio.
Args:
img (~numpy.ndarray): An array to be transformed. This is in
CHW format.
size (tuple of two ints): A tuple of two elements:
:obj:`height, width`. The size of the image after resizing.
fill (float, tuple or ~numpy.ndarray): The value of padded pixels.
If it is :class:`numpy.ndarray`,
its shape should be :math:`(C, 1, 1)`,
where :math:`C` is the number of channels of :obj:`img`.
return_param (bool): Returns information of resizing and offsetting.
Returns:
~numpy.ndarray or (~numpy.ndarray, dict):
If :obj:`return_param = False`,
returns an array :obj:`out_img` that is the result of resizing.
If :obj:`return_param = True`,
returns a tuple whose elements are :obj:`out_img, param`.
:obj:`param` is a dictionary of intermediate parameters whose
contents are listed below with key, value-type and the description
of the value.
* **y_offset** (*int*): The y coodinate of the top left corner of\
the image after placing on the canvas.
* **x_offset** (*int*): The x coordinate of the top left corner\
of the image after placing on the canvas.
* **scaled_size** (*tuple*): The size to which the image is scaled\
to before placing it on a canvas. This is a tuple of two elements:\
:obj:`height, width`.
"""
C, H, W = img.shape
out_H, out_W = size
scale_h = out_H / H
scale_w = out_W / W
scale = min(min(scale_h, scale_w), 1)
scaled_size = (int(H * scale), int(W * scale))
if scale < 1:
img = resize(img, scaled_size, interpolation=interpolation)
y_slice, x_slice = _get_pad_slice(img, size=size)
out_img = np.empty((C, out_H, out_W), dtype=img.dtype)
out_img[:] = np.array(fill).reshape((-1, 1, 1))
out_img[:, y_slice, x_slice] = img
if return_param:
param = {'y_offset': y_slice.start, 'x_offset': x_slice.start,
'scaled_size': scaled_size}
return out_img, param
else:
return out_img