def __getitem__(self, index):
# ===== fetch data ===== #
img_path = osp.join(self.data_file_dir, self.image_list[index])
img, img_align, mid = self._fetch_data(img_path)
# ===== data aug ===== #
if self.transform.resize:
img, _ = resize(img, mid, self.transform.resize)
img_align, _ = resize(img_align, mid, self.transform.resize)
if self.transform.rotate:
img, _ = random_rotate(img, mid, self.transform.rotate)
if self.transform.flip:
img, _ = random_mirror(img, mid)
if self.transform.gaussian_kernels:
img = random_gaussian_blur(img, self.transform.gaussian_kernels)
if self.transform.brightness:
img = random_brightness(img, self.transform.brightness)
img = normalize(img)
img_align = normalize(img_align)
# ===== torch data format ===== #
img = torch.FloatTensor(img.transpose(2, 0, 1))
img_align = torch.FloatTensor(img_align.transpose(2, 0, 1))
if self.stage == 'test':
return img, img_align, self.image_list[index]
else:
return img, img_align
def __call__(self, img, gt):
img, gt = random_mirror(img, gt)
if config.train_scale_array is not None:
img, gt, scale = random_scale(img, gt, config.train_scale_array)
img = normalize(img, self.img_mean, self.img_std)
# pytorch input need image value range from 0-1
crop_size = (config.image_height, config.image_width)
crop_pos = generate_random_crop_pos(img.shape[:2], crop_size)
# get legal crop start position
p_img , _ = random_crop_pad_to_shape(img, crop_pos, crop_size, 0)
p_gt , _ = random_crop_pad_to_shape(gt, crop_pos, crop_size, 255)
#gt padding value is 225, image padding value is 0
# add padding to those image with smaller shape
return p_img, p_gt
def __call__ (self, img, gt):
# can be not 320 when evaluate
img = normalize(img, self.img_mean, self.img_std)
im_original_shape = img.shape[:2] # [500 375]
print(im_original_shape)
img = mic.imresize(img, self.target_size/ float(max(im_original_shape)))
resized_shape = img.shape[:2] # 320 240
margin = [(self.target_size - resized_shape[0])//2, (self.target_size - resized_shape[1])//2 ] # 0, 40
img = cv2.copyMakeBorder(img, margin[0], self.target_size - resized_shape[0] - margin[0],
margin[1], self.target_size - resized_shape[1] - margin[1], cv2.BORDER_REFLECT_101)
assert (img.shape[0] == img.shape[1] == self.target_size)
return img, gt, dict(margin = margin, resized_shape = resized_shape, im_original_shape = im_original_shape)
def __call__ (self, img, gt, sal):
img_ = normalize(img, self.img_mean, self.img_std)
return img_, gt, sal