def predict_homography(inputs, network_id='cvgghmg', reuse=None,
is_training=True, scope='hier_hmg'):
"""Estimates homography using a selected deep neural network.
Args:
inputs: batch of input image pairs of data type float32 and of shape
[batch_size, height, width, None]
network_id: deep neural network method
reuse: whether to reuse this network weights
is_training: whether used for training or testing
scope: the scope of variables in this function
Raises:
ValueError: The nework_id was not good.
Returns:
a list of homographies at each level and a list of images warped by
the list of corresponding homographies
"""
with slim.arg_scope(models.homography_arg_scope(
weight_decay=FLAGS.weight_decay)):
if network_id == 'hier':
return models.hier_homography_estimator(
inputs, num_param=8, num_layer=FLAGS.num_layer,
num_level=FLAGS.num_level,
dropout_keep_prob=FLAGS.dropout_keep,
is_training=is_training, reuse=reuse, scope=scope)
elif network_id == 'fmask_sem':
return models.hier_homography_fmask_estimator(
inputs, num_param=8, num_layer=FLAGS.num_layer,
num_level=FLAGS.num_level,
dropout_keep_prob=FLAGS.dropout_keep,
is_training=is_training, reuse=reuse, scope=scope)
else:
raise ValueError('Unknown network_id: %s' % network_id)
def run_test():
"""Estimates the homography between two input images.
"""
image1 = cv2.imread(FLAGS.image1)
image2 = cv2.imread(FLAGS.image2)
image_list = [image1, image2]
image_norm_list = []
for i in range(2):
if FLAGS.network_id == 'fmask_sem':
image_scale = cv2.resize(image_list[i],
(FLAGS.train_width, FLAGS.train_height),
cv2.INTER_LANCZOS4)
else:
image_gray = cv2.cvtColor(image_list[i], cv2.COLOR_BGR2GRAY)
image_scale = cv2.resize(image_gray,
(FLAGS.train_width, FLAGS.train_height),
cv2.INTER_LANCZOS4)
image_norm = image_scale / 256.0 - 0.5
image_norm_list.append(image_norm)
if FLAGS.network_id == 'fmask_sem':
norm_image_pair = np.expand_dims(np.concatenate(image_norm_list, 2), axis=0)
num_channel = 3
else:
norm_image_pair = np.expand_dims(np.stack(image_norm_list, -1), axis=0)
num_channel = 1
batch_pairs = tf.placeholder(tf.float32,
[1, FLAGS.train_height, FLAGS.train_width,
2 * num_channel])
with slim.arg_scope(models.homography_arg_scope()):
if FLAGS.network_id == 'fmask_sem':
batch_hmg_prediction, _ = models.hier_homography_fmask_estimator(
batch_pairs, num_param=8, num_layer=FLAGS.num_layer,
num_level=FLAGS.num_level, is_training=False)
else:
batch_hmg_prediction, _ = models.hier_homography_estimator(
batch_pairs, num_param=8, num_layer=FLAGS.num_layer,
num_level=FLAGS.num_level, is_training=False)
batch_warped_result, _ = hmg_util.homography_warp_per_batch(
batch_pairs[Ellipsis, 0 : num_channel],
batch_hmg_prediction[FLAGS.num_level - 1])
saver = tf.Saver()
with tf.Session() as sess:
saver.restore(sess, FLAGS.model_path)
image_warp, homography_list = sess.run(
[batch_warped_result, batch_hmg_prediction],
feed_dict={batch_pairs: norm_image_pair})
for i in range(8):
logging.info('%f ', homography_list[FLAGS.num_level - 1][0][i])
cv2.imwrite('%s/input0.jpg' % FLAGS.out_dir,
(image_norm_list[0] + 0.5) * 256)
cv2.imwrite('%s/input1.jpg' % FLAGS.out_dir,
(image_norm_list[1] + 0.5) * 256)
cv2.imwrite('%s/result.jpg' % FLAGS.out_dir, (image_warp[0] + 0.5) * 256)