def scale_process(self, scale_im):
h, w = scale_im.shape[:2]
im_size_min = min(h, w)
im_size_max = max(h, w)
if im_size_max <= self.crop_size:
input_im = T.padding_im(scale_im,
target_size=(self.crop_size,
self.crop_size),
borderType=cv2.BORDER_CONSTANT)
output = self.inference(np.asarray([input_im], dtype=np.float32))
score = output[0].transpose(1, 2, 0)[:h, :w, :]
else:
stride = np.ceil(self.crop_size * self.stride_rate)
pad_im = scale_im
if im_size_min < self.crop_size:
pad_im = T.padding_im(scale_im,
target_size=(self.crop_size,
self.crop_size),
borderType=cv2.BORDER_CONSTANT)
ph, pw = pad_im.shape[:2]
h_grid = int(np.ceil(float(ph - self.crop_size) / stride)) + 1
w_grid = int(np.ceil(float(pw - self.crop_size) / stride)) + 1
data_scale = np.zeros((ph, pw, self.class_num), dtype=np.float32)
count_scale = np.zeros((ph, pw, self.class_num), dtype=np.float32)
for grid_yidx in xrange(0, h_grid):
for grid_xidx in xrange(0, w_grid):
s_x = int(grid_xidx * stride)
s_y = int(grid_yidx * stride)
e_x = min(s_x + self.crop_size, pw)
e_y = min(s_y + self.crop_size, ph)
s_x = int(e_x - self.crop_size)
s_y = int(e_y - self.crop_size)
sub_im = pad_im[s_y:e_y, s_x:e_x, :]
count_scale[s_y:e_y, s_x:e_x, :] += 1.0
input_im = T.padding_im(sub_im,
target_size=(self.crop_size,
self.crop_size),
borderType=cv2.BORDER_CONSTANT)
output = self.inference(
np.asarray([input_im], dtype=np.float32))
data_scale[s_y:e_y,
s_x:e_x, :] += output[0].transpose(1, 2, 0)
score = data_scale / count_scale
score = score[:h, :w, :]
return score
def seg_im(self, im):
""" Ignore self.scales; """
im = im.astype(np.float32, copy=True)
h, w = im.shape[:2]
normalized_im = T.normalize(im, mean=self.mean, std=self.std)
scale_im, scale_ratio = T.scale_by_max(normalized_im, long_size=self.crop_size)
input_im = T.padding_im(scale_im, target_size=(self.crop_size, self.crop_size),
borderType=cv2.BORDER_CONSTANT)
output = self.inference(np.asarray([input_im], dtype=np.float32))
score = output[0].transpose(1, 2, 0)
score_map = cv2.resize(score, None, None, fx=1. / scale_ratio, fy=1. / scale_ratio)[:h, :w, :]
return score_map.argmax(2)
def seg_im(self, im):
""" Ignore self.scales; """
im = im.astype(np.float32, copy=True)
h, w = im.shape[:2]
normalized_im = T.normalize(im, mean=self.mean, std=self.std)
scale_im, scale_ratio = T.scale_by_max(normalized_im,
long_size=self.crop_size)
input_im = T.padding_im(scale_im,
target_size=(self.crop_size, self.crop_size),
borderType=cv2.BORDER_CONSTANT,
borderValue=(0.0, 0.0, 0.0))
score = self.caffe_process(input_im)
score_map = cv2.resize(score,
None,
None,
fx=1. / scale_ratio,
fy=1. / scale_ratio)[:h, :w, :]
return score_map.argmax(2)