def __call__(self, sample):
_target = sample[self.mask_elem]
for elem in self.crop_elems:
_img = sample[elem]
if self.mask_elem == elem:
_tmp_img = _img
_tmp_target = _target
if np.max(_target) == 0:
_crop = np.zeros(_tmp_img.shape, dtype=_img.dtype)
else:
_crop = helpers.crop_from_mask(_tmp_img,
_tmp_target,
relax=self.relax,
zero_pad=self.zero_pad)
else:
if np.max(_target) == 0:
_crop = np.zeros(_img.shape, dtype=_img.dtype)
else:
_tmp_target = _target
_crop = helpers.crop_from_mask(_img,
_tmp_target,
relax=self.relax,
zero_pad=self.zero_pad)
sample['crop_' + elem] = _crop
return sample
def __call__(self, sample):
_target = sample[self.mask_elem]
if _target.ndim == 2:
_target = np.expand_dims(_target, axis=-1)
for elem in self.crop_elems:
_img = sample[elem]
_crop = []
if self.mask_elem == elem:
if _img.ndim == 2:
_img = np.expand_dims(_img, axis=-1)
for k in range(0, _target.shape[-1]):
_tmp_img = _img[..., k]
_tmp_target = _target[..., k]
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_tmp_img.shape, dtype=_img.dtype))
else:
_crop.append(helpers.crop_from_mask(_tmp_img, _tmp_target, relax=self.relax, zero_pad=self.zero_pad))
else:
for k in range(0, _target.shape[-1]):
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_img.shape, dtype=_img.dtype))
else:
_tmp_target = _target[..., k]
_crop.append(helpers.crop_from_mask(_img, _tmp_target, relax=self.relax, zero_pad=self.zero_pad))
if len(_crop) == 1:
sample['crop_' + elem] = _crop[0]
else:
sample['crop_' + elem] = _crop
return sample
def __call__(self, sample):
self.dz = random.randint(350, 400)
_target = sample[self.mask_elem]
if len(np.unique(_target)) == 1:
sample['crop_image'] = sample['image']
sample['crop_gt'] = sample['gt']
return sample
if _target.ndim == 2:
_target = np.expand_dims(_target, axis=-1)
for elem in self.crop_elems:
_img = sample[elem]
_crop = []
### dynamic relax crop ###
bbox = helpers.get_bbox(_target)
d = np.maximum(bbox[2] - bbox[0], bbox[3] - bbox[1])
sample['temp'] = d
zoom_factor = self.dz / d
crop_relax = (512 - d * zoom_factor) / (2 * zoom_factor)
self.crop_relax = np.ceil(crop_relax).astype(int)
self.crop_relax = np.maximum(15, self.crop_relax)
sample['crop_relax'] = self.crop_relax
### ###
if self.mask_elem == elem:
if _img.ndim == 2:
_img = np.expand_dims(_img, axis=-1)
for k in range(0, _target.shape[-1]):
_tmp_img = _img[..., k]
_tmp_target = _target[..., k]
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_tmp_img.shape,
dtype=_img.dtype))
else:
_crop.append(
helpers.crop_from_mask(_tmp_img,
_tmp_target,
relax=self.crop_relax,
zero_pad=self.zero_pad))
else:
for k in range(0, _target.shape[-1]):
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_img.shape, dtype=_img.dtype))
else:
_tmp_target = _target[..., k]
_crop.append(
helpers.crop_from_mask(_img,
_tmp_target,
relax=self.crop_relax,
zero_pad=self.zero_pad))
if len(_crop) == 1:
sample['crop_' + elem] = _crop[0]
else:
sample['crop_' + elem] = _crop
return sample
def __call__(self, sample):
if self.dummy:
for elem in self.crop_elems:
sample['crop_' + elem] = sample[elem].copy()
return sample
_target = sample[self.mask_elem]
if _target.ndim == 2:
_target = np.expand_dims(_target, axis=-1)
for elem in self.crop_elems:
_img = sample[elem]
_crop = []
if self.mask_elem == elem:
if _img.ndim == 2:
_img = np.expand_dims(_img, axis=-1)
for k in range(0, _target.shape[-1]):
_tmp_img = _img[..., k]
_tmp_target = _target[..., k]
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_tmp_img.shape,
dtype=_img.dtype))
else:
_crop.append(
np.ascontiguousarray(
helpers.crop_from_mask(
_tmp_img,
_tmp_target,
relax=self.relax,
zero_pad=self.zero_pad,
for_grid=self.for_grid,
grid_size=self.grid_size)))
else:
for k in range(0, _target.shape[-1]):
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_img.shape, dtype=_img.dtype))
else:
_tmp_target = _target[..., k]
_crop.append(
np.ascontiguousarray(
helpers.crop_from_mask(
_img,
_tmp_target,
relax=self.relax,
zero_pad=self.zero_pad,
for_grid=self.for_grid,
grid_size=self.grid_size)))
if len(_crop) == 1:
sample['crop_' + elem] = _crop[0]
else:
sample['crop_' + elem] = _crop
sample['crop_mask'] = sample['crop_gt']
return sample
def __call__(self, sample):
if self.is_val == False:
self.dz = np.array(np.random.randint(self.min_,self.max_)).astype(np.float32)
_target = sample[self.mask_elem]
if len(np.unique(_target)) == 1:
sample['crop_image'] = sample['image']
sample['crop_gt'] = sample['gt']
return sample
if _target.ndim == 2:
_target = np.expand_dims(_target, axis=-1)
for elem in self.crop_elems:
_img = sample[elem]
_crop = []
### dynamic relax crop ###
bbox = helpers.get_bbox(_target)
d = np.maximum(bbox[2] - bbox[0], bbox[3] - bbox[1])
if d < 1:
print("Very small objects detected")
print(sample['id'])
zoom_factor = self.dz/d
crop_relax = (self.d-d*zoom_factor)/(2*zoom_factor)
crop_relax = np.maximum(crop_relax, self.thresh)
self.crop_relax = np.ceil(crop_relax).astype(int)
sample['crop_relax'] = self.crop_relax
### ###
if self.mask_elem == elem:
if _img.ndim == 2:
_img = np.expand_dims(_img, axis=-1)
for k in range(0, _target.shape[-1]):
_tmp_img = _img[..., k]
_tmp_target = _target[..., k]
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_tmp_img.shape, dtype=_img.dtype))
else:
_crop.append(helpers.crop_from_mask(_tmp_img, _tmp_target, relax=self.crop_relax, zero_pad=self.zero_pad))
else:
for k in range(0, _target.shape[-1]):
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_img.shape, dtype=_img.dtype))
else:
_tmp_target = _target[..., k]
_crop.append(helpers.crop_from_mask(_img, _tmp_target, relax=self.crop_relax, zero_pad=self.zero_pad))
if len(_crop) == 1:
sample['crop_' + elem] = _crop[0]
else:
sample['crop_' + elem] = _crop
return sample
def __call__(self, sample):
_target = sample[self.mask_elem]
# enable adaptive calculation of bbox relaxation
if self.adaptive_relax:
mean_shape = np.mean(_target.shape[:2])
# 428 is the average size in PASCAL trainaug
relax = int(self.relax * mean_shape / 428.)
else:
relax = self.relax
sample['meta']['relax'] = relax
if _target.ndim == 2:
_target = np.expand_dims(_target, axis=-1)
for elem in self.crop_elems:
_img = sample[elem]
_crop = []
if self.mask_elem == elem:
if _img.ndim == 2:
_img = np.expand_dims(_img, axis=-1)
for k in range(0, _target.shape[-1]):
_tmp_img = _img[..., k]
_tmp_target = _target[..., k]
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_tmp_img.shape, dtype=_img.dtype))
else:
_crop.append(helpers.crop_from_mask(_tmp_img, _tmp_target, relax=relax, zero_pad=self.zero_pad))
else:
for k in range(0, _target.shape[-1]):
if np.max(_target[..., k]) == 0:
_crop.append(np.zeros(_img.shape, dtype=_img.dtype))
else:
_tmp_target = _target[..., k]
_crop.append(helpers.crop_from_mask(_img, _tmp_target, relax=relax, zero_pad=self.zero_pad))
if len(_crop) == 1:
sample[self.prefix + elem] = _crop[0]
else:
sample[self.prefix + elem] = _crop
return sample
def __call__(self, sample):
_target_list = sample[self.mask_elem]
sample['crop_image'] = []
sample['crop_gt'] = []
sample['crop_polygon'] = []
sample['valid'] = []
for i_comp, _target in enumerate(_target_list):
_target = np.expand_dims(_target, axis=-1)
elem = 'image'
_img = sample[elem]
if _img.ndim == 2:
_img = np.expand_dims(_img, axis=-1)
_crop_img = []
for k in range(0, _target.shape[-1]):
if np.max(_target[..., k]) == 0:
_crop_img.append(np.zeros(_img.shape, dtype=_img.dtype))
else:
_tmp_target = _target[..., k]
_crop_img.append(
helpers.crop_from_mask(_img,
_tmp_target,
relax=self.relax,
zero_pad=self.zero_pad))
sample['crop_image'].append(_crop_img[0])
elem = 'gt'
_img = sample[elem][i_comp]
_crop_img = []
if _img.ndim == 2:
_img = np.expand_dims(_img, axis=-1)
for k in range(0, _target.shape[-1]):
_tmp_img = _img[..., k]
_tmp_target = _target[..., k]
if np.max(_target[..., k]) == 0:
_crop_img.append(np.zeros(_tmp_img.shape,
dtype=_img.dtype))
else:
_crop_img.append(
helpers.crop_from_mask(_tmp_img,
_tmp_target,
relax=self.relax,
zero_pad=self.zero_pad))
sample['crop_gt'].append(_crop_img[0])
# import ipdb
# ipdb.set_trace()
elem = 'gt_polygon'
gt_polygon = sample[elem][i_comp]
valid = 1
try:
_crop_polygon = helpers.crop_from_mask_polygon(
gt_polygon,
_img,
_target[..., 0],
relax=self.relax,
zero_pad=self.zero_pad)
except:
_crop_polygon = np.asarray([[0.5, 0.5], [0.5, 0.6], [0.6, 0.6],
[0.6, 0.5]])
valid = 0
sample['valid'].append(valid)
sample['crop_polygon'].append(_crop_polygon)
return sample