mmedit.datasets.pipelines.crop 源代码
# Copyright (c) OpenMMLab. All rights reserved.
import mmcv
import numpy as np
from torch.nn.modules.utils import _pair
from ..registry import PIPELINES
from .utils import random_choose_unknown
[文档]@PIPELINES.register_module()
class Crop:
"""Crop data to specific size for training.
Args:
keys (Sequence[str]): The images to be cropped.
crop_size (Tuple[int]): Target spatial size (h, w).
random_crop (bool): If set to True, it will random crop
image. Otherwise, it will work as center crop.
"""
def __init__(self, keys, crop_size, random_crop=True):
if not mmcv.is_tuple_of(crop_size, int):
raise TypeError(
'Elements of crop_size must be int and crop_size must be'
f' tuple, but got {type(crop_size[0])} in {type(crop_size)}')
self.keys = keys
self.crop_size = crop_size
self.random_crop = random_crop
def _crop(self, data):
if not isinstance(data, list):
data_list = [data]
else:
data_list = data
crop_bbox_list = []
data_list_ = []
for item in data_list:
data_h, data_w = item.shape[:2]
crop_h, crop_w = self.crop_size
crop_h = min(data_h, crop_h)
crop_w = min(data_w, crop_w)
if self.random_crop:
x_offset = np.random.randint(0, data_w - crop_w + 1)
y_offset = np.random.randint(0, data_h - crop_h + 1)
else:
x_offset = max(0, (data_w - crop_w)) // 2
y_offset = max(0, (data_h - crop_h)) // 2
crop_bbox = [x_offset, y_offset, crop_w, crop_h]
item_ = item[y_offset:y_offset + crop_h,
x_offset:x_offset + crop_w, ...]
crop_bbox_list.append(crop_bbox)
data_list_.append(item_)
if not isinstance(data, list):
return data_list_[0], crop_bbox_list[0]
return data_list_, crop_bbox_list
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Returns:
dict: A dict containing the processed data and information.
"""
for k in self.keys:
data_, crop_bbox = self._crop(results[k])
results[k] = data_
results[k + '_crop_bbox'] = crop_bbox
results['crop_size'] = self.crop_size
return results
def __repr__(self):
repr_str = self.__class__.__name__
repr_str += (f'keys={self.keys}, crop_size={self.crop_size}, '
f'random_crop={self.random_crop}')
return repr_str
[文档]@PIPELINES.register_module()
class CropSequence(Crop):
"""Crop a sequence to specific size for training.
The main difference to 'Crop' is that the region to be cropped is the same
for every images in the sequence.
Args:
keys (Sequence[str]): The images to be cropped.
crop_size (Tuple[int]): Target spatial size (h, w).
random_crop (bool): If set to True, it will random crop
image. Otherwise, it will work as center crop.
"""
def _crop(self, data):
if not isinstance(data, list):
raise TypeError(f'Input must be a list, but got {type(data)}.')
# determine crop location. Must be the same for all images
data_h, data_w = data[0].shape[:2]
crop_h, crop_w = self.crop_size
crop_h = min(data_h, crop_h)
crop_w = min(data_w, crop_w)
if self.random_crop:
x_offset = np.random.randint(0, data_w - crop_w + 1)
y_offset = np.random.randint(0, data_h - crop_h + 1)
else:
x_offset = max(0, (data_w - crop_w)) // 2
y_offset = max(0, (data_h - crop_h)) // 2
crop_bbox = [x_offset, y_offset, crop_w, crop_h]
data_list = []
for item in data:
item = item[y_offset:y_offset + crop_h, x_offset:x_offset + crop_w,
...]
data_list.append(item)
return data_list, crop_bbox
[文档]@PIPELINES.register_module()
class FixedCrop:
"""Crop paired data (at a specific position) to specific size for training.
Args:
keys (Sequence[str]): The images to be cropped.
crop_size (Tuple[int]): Target spatial size (h, w).
crop_pos (Tuple[int]): Specific position (x, y). If set to None,
random initialize the position to crop paired data batch.
"""
def __init__(self, keys, crop_size, crop_pos=None):
if not mmcv.is_tuple_of(crop_size, int):
raise TypeError(
'Elements of crop_size must be int and crop_size must be'
f' tuple, but got {type(crop_size[0])} in {type(crop_size)}')
if not mmcv.is_tuple_of(crop_pos, int) and (crop_pos is not None):
raise TypeError(
'Elements of crop_pos must be int and crop_pos must be'
f' tuple or None, but got {type(crop_pos[0])} in '
f'{type(crop_pos)}')
self.keys = keys
self.crop_size = crop_size
self.crop_pos = crop_pos
def _crop(self, data, x_offset, y_offset, crop_w, crop_h):
crop_bbox = [x_offset, y_offset, crop_w, crop_h]
data_ = data[y_offset:y_offset + crop_h, x_offset:x_offset + crop_w,
...]
return data_, crop_bbox
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Returns:
dict: A dict containing the processed data and information.
"""
if isinstance(results[self.keys[0]], list):
data_h, data_w = results[self.keys[0]][0].shape[:2]
else:
data_h, data_w = results[self.keys[0]].shape[:2]
crop_h, crop_w = self.crop_size
crop_h = min(data_h, crop_h)
crop_w = min(data_w, crop_w)
if self.crop_pos is None:
x_offset = np.random.randint(0, data_w - crop_w + 1)
y_offset = np.random.randint(0, data_h - crop_h + 1)
else:
x_offset, y_offset = self.crop_pos
crop_w = min(data_w - x_offset, crop_w)
crop_h = min(data_h - y_offset, crop_h)
for k in self.keys:
images = results[k]
is_list = isinstance(images, list)
if not is_list:
images = [images]
cropped_images = []
crop_bbox = None
for image in images:
# In fixed crop for paired images, sizes should be the same
if (image.shape[0] != data_h or image.shape[1] != data_w):
raise ValueError(
'The sizes of paired images should be the same. '
f'Expected ({data_h}, {data_w}), '
f'but got ({image.shape[0]}, '
f'{image.shape[1]}).')
data_, crop_bbox = self._crop(image, x_offset, y_offset,
crop_w, crop_h)
cropped_images.append(data_)
results[k + '_crop_bbox'] = crop_bbox
if not is_list:
cropped_images = cropped_images[0]
results[k] = cropped_images
results['crop_size'] = self.crop_size
results['crop_pos'] = self.crop_pos
return results
def __repr__(self):
repr_str = self.__class__.__name__
repr_str += (f'keys={self.keys}, crop_size={self.crop_size}, '
f'crop_pos={self.crop_pos}')
return repr_str
[文档]@PIPELINES.register_module()
class PairedRandomCrop:
"""Paried random crop.
It crops a pair of lq and gt images with corresponding locations.
It also supports accepting lq list and gt list.
Required keys are "scale", "lq", and "gt",
added or modified keys are "lq" and "gt".
Args:
gt_patch_size (int): cropped gt patch size.
"""
def __init__(self, gt_patch_size):
self.gt_patch_size = gt_patch_size
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Returns:
dict: A dict containing the processed data and information.
"""
scale = results['scale']
lq_patch_size = self.gt_patch_size // scale
lq_is_list = isinstance(results['lq'], list)
if not lq_is_list:
results['lq'] = [results['lq']]
gt_is_list = isinstance(results['gt'], list)
if not gt_is_list:
results['gt'] = [results['gt']]
h_lq, w_lq, _ = results['lq'][0].shape
h_gt, w_gt, _ = results['gt'][0].shape
if h_gt != h_lq * scale or w_gt != w_lq * scale:
raise ValueError(
f'Scale mismatches. GT ({h_gt}, {w_gt}) is not {scale}x ',
f'multiplication of LQ ({h_lq}, {w_lq}).')
if h_lq < lq_patch_size or w_lq < lq_patch_size:
raise ValueError(
f'LQ ({h_lq}, {w_lq}) is smaller than patch size ',
f'({lq_patch_size}, {lq_patch_size}). Please check '
f'{results["lq_path"][0]} and {results["gt_path"][0]}.')
# randomly choose top and left coordinates for lq patch
top = np.random.randint(h_lq - lq_patch_size + 1)
left = np.random.randint(w_lq - lq_patch_size + 1)
# crop lq patch
results['lq'] = [
v[top:top + lq_patch_size, left:left + lq_patch_size, ...]
for v in results['lq']
]
# crop corresponding gt patch
top_gt, left_gt = int(top * scale), int(left * scale)
results['gt'] = [
v[top_gt:top_gt + self.gt_patch_size,
left_gt:left_gt + self.gt_patch_size, ...] for v in results['gt']
]
if not lq_is_list:
results['lq'] = results['lq'][0]
if not gt_is_list:
results['gt'] = results['gt'][0]
return results
def __repr__(self):
repr_str = self.__class__.__name__
repr_str += f'(gt_patch_size={self.gt_patch_size})'
return repr_str
[文档]@PIPELINES.register_module()
class CropAroundCenter:
"""Randomly crop the images around unknown area in the center 1/4 images.
This cropping strategy is adopted in GCA matting. The `unknown area` is the
same as `semi-transparent area`.
https://arxiv.org/pdf/2001.04069.pdf
It retains the center 1/4 images and resizes the images to 'crop_size'.
Required keys are "fg", "bg", "trimap" and "alpha", added or modified keys
are "crop_bbox", "fg", "bg", "trimap" and "alpha".
Args:
crop_size (int | tuple): Desired output size. If int, square crop is
applied.
"""
def __init__(self, crop_size):
if mmcv.is_tuple_of(crop_size, int):
assert len(crop_size) == 2, 'length of crop_size must be 2.'
elif not isinstance(crop_size, int):
raise TypeError('crop_size must be int or a tuple of int, but got '
f'{type(crop_size)}')
self.crop_size = _pair(crop_size)
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Returns:
dict: A dict containing the processed data and information.
"""
fg = results['fg']
alpha = results['alpha']
trimap = results['trimap']
bg = results['bg']
h, w = fg.shape[:2]
assert bg.shape == fg.shape, (f'shape of bg {bg.shape} should be the '
f'same as fg {fg.shape}.')
crop_h, crop_w = self.crop_size
# Make sure h >= crop_h, w >= crop_w. If not, rescale imgs
rescale_ratio = max(crop_h / h, crop_w / w)
if rescale_ratio > 1:
new_h = max(int(h * rescale_ratio), crop_h)
new_w = max(int(w * rescale_ratio), crop_w)
fg = mmcv.imresize(fg, (new_w, new_h), interpolation='nearest')
alpha = mmcv.imresize(
alpha, (new_w, new_h), interpolation='nearest')
trimap = mmcv.imresize(
trimap, (new_w, new_h), interpolation='nearest')
bg = mmcv.imresize(bg, (new_w, new_h), interpolation='bicubic')
h, w = new_h, new_w
# resize to 1/4 to ignore small unknown patches
small_trimap = mmcv.imresize(
trimap, (w // 4, h // 4), interpolation='nearest')
# find unknown area in center 1/4 region
margin_h, margin_w = crop_h // 2, crop_w // 2
sample_area = small_trimap[margin_h // 4:(h - margin_h) // 4,
margin_w // 4:(w - margin_w) // 4]
unknown_xs, unknown_ys = np.where(sample_area == 128)
unknown_num = len(unknown_xs)
if unknown_num < 10:
# too few unknown area in the center, crop from the whole image
top = np.random.randint(0, h - crop_h + 1)
left = np.random.randint(0, w - crop_w + 1)
else:
idx = np.random.randint(unknown_num)
top = unknown_xs[idx] * 4
left = unknown_ys[idx] * 4
bottom = top + crop_h
right = left + crop_w
results['fg'] = fg[top:bottom, left:right]
results['alpha'] = alpha[top:bottom, left:right]
results['trimap'] = trimap[top:bottom, left:right]
results['bg'] = bg[top:bottom, left:right]
results['crop_bbox'] = (left, top, right, bottom)
return results
def __repr__(self):
return self.__class__.__name__ + f'(crop_size={self.crop_size})'
[文档]@PIPELINES.register_module()
class CropAroundUnknown:
"""Crop around unknown area with a randomly selected scale.
Randomly select the w and h from a list of (w, h).
Required keys are the keys in argument `keys`, added or
modified keys are "crop_bbox" and the keys in argument `keys`.
This class assumes value of "alpha" ranges from 0 to 255.
Args:
keys (Sequence[str]): The images to be cropped. It must contain
'alpha'. If unknown_source is set to 'trimap', then it must also
contain 'trimap'.
crop_sizes (list[int | tuple[int]]): List of (w, h) to be selected.
unknown_source (str, optional): Unknown area to select from. It must be
'alpha' or 'tirmap'. Default to 'alpha'.
interpolations (str | list[str], optional): Interpolation method of
mmcv.imresize. The interpolation operation will be applied when
image size is smaller than the crop_size. If given as a list of
str, it should have the same length as `keys`. Or if given as a
str all the keys will be resized with the same method.
Default to 'bilinear'.
"""
def __init__(self,
keys,
crop_sizes,
unknown_source='alpha',
interpolations='bilinear'):
if 'alpha' not in keys:
raise ValueError(f'"alpha" must be in keys, but got {keys}')
self.keys = keys
if not isinstance(crop_sizes, list):
raise TypeError(
f'Crop sizes must be list, but got {type(crop_sizes)}.')
self.crop_sizes = [_pair(crop_size) for crop_size in crop_sizes]
if not mmcv.is_tuple_of(self.crop_sizes[0], int):
raise TypeError('Elements of crop_sizes must be int or tuple of '
f'int, but got {type(self.crop_sizes[0][0])}.')
if unknown_source not in ['alpha', 'trimap']:
raise ValueError('unknown_source must be "alpha" or "trimap", '
f'but got {unknown_source}')
if unknown_source not in keys:
# it could only be trimap, since alpha is checked before
raise ValueError(
'if unknown_source is "trimap", it must also be set in keys')
self.unknown_source = unknown_source
if isinstance(interpolations, str):
self.interpolations = [interpolations] * len(self.keys)
elif mmcv.is_list_of(interpolations,
str) and len(interpolations) == len(self.keys):
self.interpolations = interpolations
else:
raise TypeError(
'interpolations must be a str or list of str with '
f'the same length as keys, but got {interpolations}')
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Returns:
dict: A dict containing the processed data and information.
"""
h, w = results[self.keys[0]].shape[:2]
rand_ind = np.random.randint(len(self.crop_sizes))
crop_h, crop_w = self.crop_sizes[rand_ind]
# Make sure h >= crop_h, w >= crop_w. If not, rescale imgs
rescale_ratio = max(crop_h / h, crop_w / w)
if rescale_ratio > 1:
h = max(int(h * rescale_ratio), crop_h)
w = max(int(w * rescale_ratio), crop_w)
for key, interpolation in zip(self.keys, self.interpolations):
results[key] = mmcv.imresize(
results[key], (w, h), interpolation=interpolation)
# Select the cropping top-left point which is an unknown pixel
if self.unknown_source == 'alpha':
unknown = (results['alpha'] > 0) & (results['alpha'] < 255)
else:
unknown = results['trimap'] == 128
top, left = random_choose_unknown(unknown.squeeze(), (crop_h, crop_w))
bottom = top + crop_h
right = left + crop_w
for key in self.keys:
results[key] = results[key][top:bottom, left:right]
results['crop_bbox'] = (left, top, right, bottom)
return results
def __repr__(self):
repr_str = self.__class__.__name__
repr_str += (f'(keys={self.keys}, crop_sizes={self.crop_sizes}, '
f"unknown_source='{self.unknown_source}', "
f'interpolations={self.interpolations})')
return repr_str
[文档]@PIPELINES.register_module()
class CropAroundFg:
"""Crop around the whole foreground in the segmentation mask.
Required keys are "seg" and the keys in argument `keys`.
Meanwhile, "seg" must be in argument `keys`. Added or modified keys are
"crop_bbox" and the keys in argument `keys`.
Args:
keys (Sequence[str]): The images to be cropped. It must contain
'seg'.
bd_ratio_range (tuple, optional): The range of the boundary (bd) ratio
to select from. The boundary ratio is the ratio of the boundary to
the minimal bbox that contains the whole foreground given by
segmentation. Default to (0.1, 0.4).
test_mode (bool): Whether use test mode. In test mode, the tight crop
area of foreground will be extended to the a square.
Default to False.
"""
def __init__(self, keys, bd_ratio_range=(0.1, 0.4), test_mode=False):
if 'seg' not in keys:
raise ValueError(f'"seg" must be in keys, but got {keys}')
if (not mmcv.is_tuple_of(bd_ratio_range, float)
or len(bd_ratio_range) != 2):
raise TypeError('bd_ratio_range must be a tuple of 2 int, but got '
f'{bd_ratio_range}')
self.keys = keys
self.bd_ratio_range = bd_ratio_range
self.test_mode = test_mode
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Returns:
dict: A dict containing the processed data and information.
"""
seg = results['seg']
height, width = seg.shape[:2]
# get foreground bbox
fg_coor = np.array(np.where(seg))
top, left = np.amin(fg_coor, axis=1)
bottom, right = np.amax(fg_coor, axis=1)
# enlarge bbox
long_side = np.maximum(bottom - top, right - left)
if self.test_mode:
bottom = top + long_side
right = left + long_side
boundary_ratio = np.random.uniform(*self.bd_ratio_range)
boundary = int(np.round(boundary_ratio * long_side))
# NOTE: Different from the original repo, we keep track of the four
# corners of the bbox (left, top, right, bottom) while the original
# repo use (top, left, height, width) to represent bbox. This may
# introduce an difference of 1 pixel.
top = max(top - boundary, 0)
left = max(left - boundary, 0)
bottom = min(bottom + boundary, height)
right = min(right + boundary, width)
for key in self.keys:
results[key] = results[key][top:bottom, left:right]
results['crop_bbox'] = (left, top, right, bottom)
return results
[文档]@PIPELINES.register_module()
class ModCrop:
"""Mod crop gt images, used during testing.
Required keys are "scale" and "gt",
added or modified keys are "gt".
"""
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Returns:
dict: A dict containing the processed data and information.
"""
img = results['gt'].copy()
scale = results['scale']
if img.ndim in [2, 3]:
h, w = img.shape[0], img.shape[1]
h_remainder, w_remainder = h % scale, w % scale
img = img[:h - h_remainder, :w - w_remainder, ...]
else:
raise ValueError(f'Wrong img ndim: {img.ndim}.')
results['gt'] = img
return results
[文档]@PIPELINES.register_module()
class CropLike:
"""Crop/pad the image in the target_key according to the size of image
in the reference_key .
Args:
target_key (str): The key needs to be cropped.
reference_key (str | None): The reference key, need its size.
Default: None.
"""
def __init__(self, target_key, reference_key=None):
assert reference_key and target_key
self.target_key = target_key
self.reference_key = reference_key
def __call__(self, results):
"""Call function.
Args:
results (dict): A dict containing the necessary information and
data for augmentation.
Require self.target_key and self.reference_key.
Returns:
dict: A dict containing the processed data and information.
Modify self.target_key.
"""
size = results[self.reference_key].shape
old_image = results[self.target_key]
old_size = old_image.shape
h, w = old_size[:2]
new_size = size[:2] + old_size[2:]
h_cover, w_cover = min(h, size[0]), min(w, size[1])
format_image = np.zeros(new_size, dtype=old_image.dtype)
format_image[:h_cover, :w_cover] = old_image[:h_cover, :w_cover]
results[self.target_key] = format_image
return results
def __repr__(self):
return (self.__class__.__name__ + f' target_key={self.target_key}, ' +
f'reference_key={self.reference_key}')