def test_gaussian_blur(self):
num_blur_steps = 1
mask = np.zeros(shape=(1, 3, 3, 3))
mask[:, 1, 1, :] = 1
with self.test_session() as s:
blurred_mask = s.run(gaussian_utils.blur_mask(mask, num_blur_steps))
self.assertAllLess(blurred_mask[:, 1, 1, :], 1)
def confidence_reconstruction_loss(y_true, y_pred, mask, num_steps,
gaussian_kernel_size, gaussian_kernel_std):
mask_blurred = gaussian_utils.blur_mask(mask, num_steps,
gaussian_kernel_size,
gaussian_kernel_std)
valid_mask = 1 - mask
diff = K.abs(y_true - y_pred)
l1 = K.mean(diff * valid_mask + diff * mask_blurred, axis=[1, 2, 3])
return l1
def test_rectangle_mask_gaussian_blur_1_conv_step(self):
num_conv_steps = [1, 2, 3, 4, 5, 10]
mask = np.zeros(shape=(512, 512, 3))
mask[128:384, 128:384, :] = 1
mask = np.expand_dims(mask, 0)
with self.test_session() as s:
for steps in num_conv_steps:
blurred_mask = s.run(gaussian_utils.blur_mask(mask, steps))
blurred_mask = blurred_mask[0]
blurred_mask = cv2.resize(blurred_mask, (128, 128), interpolation=cv2.INTER_CUBIC)
cv2.imwrite('./test_results/rectangle_blurred_mask_{}_step.png'.format(steps),
blurred_mask * 255)
def test_small_nvidia_mask_gaussian_blur(self):
num_conv_steps = [1, 2, 3, 4, 5, 10]
mask = cv2.imread('./pics/small_mask.png')
mask[mask == 255] = 1
mask = np.expand_dims(mask, 0)
with self.test_session() as s:
for steps in num_conv_steps:
blurred_mask = s.run(gaussian_utils.blur_mask(mask, steps))
blurred_mask = blurred_mask[0]
blurred_mask = cv2.resize(blurred_mask, (128, 128), interpolation=cv2.INTER_CUBIC)
cv2.imwrite('./test_results/small_blurred_mask_{}_step.png'.format(steps),
blurred_mask * 255)