def _cumsum_gradient(score_map, box):
b, n, m, c = utils.get_tensor_shape(score_map)
_, p, _ = utils.get_tensor_shape(box)
expanded_box = _get_expanded_box(box,
img_h=n,
img_w=m,
border_ratio=border_ratio)
(box_h, box_w) = _get_box_shape(box)
(expanded_box_h, expanded_box_w) = _get_box_shape(expanded_box)
cumsum = imgproc.calc_cumsum_2d(
score_map, tf.concat([box, expanded_box], axis=1))
area = tf.expand_dims(tf.cast(box_h * box_w, tf.float32), axis=-1)
area_border = tf.expand_dims(tf.cast(
expanded_box_h * expanded_box_w - box_h * box_w, tf.float32),
axis=-1)
avg_val = tf.div(cumsum[:, :p, :], tf.maximum(_SMALL_NUMBER, area))
avg_val_in_border = tf.div(cumsum[:, p:, :] - cumsum[:, :p, :],
tf.maximum(_SMALL_NUMBER, area_border))
return avg_val - avg_val_in_border
def test_calc_cumsum_2d(self):
tf.reset_default_graph()
image = tf.placeholder(tf.float32, shape=[None, None, None])
box = tf.placeholder(tf.int64, shape=[None, 4])
cumsum = imgproc.calc_cumsum_2d(tf.expand_dims(image, axis=0),
tf.expand_dims(box, axis=0))[0]
with self.test_session() as sess:
# 3x3 image, 1 channel.
result = sess.run(cumsum,
feed_dict={
image: [[[1], [2], [3]], [[4], [5], [6]],
[[7], [8], [9]]],
box: [
[0, 0, 1, 1],
[2, 2, 3, 3],
[1, 1, 2, 2],
[0, 0, 2, 2],
[1, 1, 3, 3],
[0, 0, 3, 3],
[0, 0, 2, 3],
]
})
self.assertAllClose(result,
[[1], [9], [5], [12], [28], [45], [21]])
# 3x3 image, 2 channels.
result = sess.run(cumsum,
feed_dict={
image: [[[1, 1], [2, 2], [3, 3]],
[[4, 1], [5, 2], [6, 3]],
[[7, 1], [8, 2], [9, 3]]],
box: [
[0, 0, 1, 1],
[2, 2, 3, 3],
[1, 1, 2, 2],
[0, 0, 2, 2],
[1, 1, 3, 3],
[0, 0, 3, 3],
[0, 0, 2, 3],
]
})
self.assertAllClose(result, [[1, 1], [9, 3], [5, 2], [12, 6],
[28, 10], [45, 18], [21, 12]])
def _cumsum_gradient(score_map, box):
b, n, m, c = utils.get_tensor_shape(score_map)
_, p, _ = utils.get_tensor_shape(box)
# Leave a border for the image border.
ymin, xmin, ymax, xmax = tf.unstack(box, axis=-1)
ymin, xmin = tf.maximum(ymin, 2), tf.maximum(xmin, 2)
ymax, xmax = tf.minimum(ymax, tf.to_int64(n - 2)), tf.minimum(
xmax, tf.to_int64(m - 2))
box = tf.stack([ymin, xmin, ymax, xmax], axis=-1)
box_exp = _get_expanded_box(box,
img_h=n,
img_w=m,
border_ratio=border_ratio)
box_list = [box, box_exp]
area_list = [
tf.cast(_get_box_area(b), tf.float32) for b in box_list
]
cumsum = imgproc.calc_cumsum_2d(score_map,
tf.concat(box_list, axis=1))
cumsum_list = [
cumsum[:, i * p:(i + 1) * p, :] for i in range(len(box_list))
]
# The main box has to be valid, including the four shrinked boxes.
assert_op = tf.Assert(
tf.reduce_all(tf.greater(area_list[0], 0)),
["Check area of the main box failed:", area_list[0]])
with tf.control_dependencies([assert_op]):
border_area = area_list[1] - area_list[0]
border_cumsum = cumsum_list[1] - cumsum_list[0]
border_avg = tf.div(
border_cumsum,
tf.maximum(_SMALL_NUMBER,
tf.expand_dims(border_area, axis=-1)))
box_avg = tf.div(
cumsum_list[0],
tf.maximum(_SMALL_NUMBER,
tf.expand_dims(area_list[0], axis=-1)))
return purity_weight * box_avg - border_avg
def _cumsum_gradient(score_map, box):
b, n, m, c = utils.get_tensor_shape(score_map)
_, p, _ = utils.get_tensor_shape(box)
n, m = tf.cast(n, tf.int64), tf.cast(m, tf.int64)
# Leave a border for the image border.
ymin, xmin, ymax, xmax = tf.unstack(box, axis=-1)
ymin, xmin = tf.maximum(ymin, 3), tf.maximum(xmin, 3)
ymax, xmax = tf.minimum(ymax, n - 3), tf.minimum(xmax, m - 3)
box = tf.stack([ymin, xmin, ymax, xmax], axis=-1)
box_exp = _get_expanded_box(box,
img_h=n,
img_w=m,
border_ratio=border_ratio)
box_shr = _get_shrinked_box(box,
img_h=n,
img_w=m,
border_ratio=border_ratio)
ymin, xmin, ymax, xmax = tf.unstack(box, axis=-1)
ymin_exp, xmin_exp, ymax_exp, xmax_exp = tf.unstack(box_exp,
axis=-1)
ymin_shr, xmin_shr, ymax_shr, xmax_shr = tf.unstack(box_shr,
axis=-1)
box_list = [
box,
tf.stack([ymin, xmin, ymax, xmin_shr], axis=-1),
tf.stack([ymin, xmax_shr, ymax, xmax], axis=-1),
tf.stack([ymin, xmin, ymin_shr, xmax], axis=-1),
tf.stack([ymax_shr, xmin, ymax, xmax], axis=-1),
tf.stack([ymin, xmin_exp, ymax, xmin], axis=-1),
tf.stack([ymin, xmax, ymax, xmax_exp], axis=-1),
tf.stack([ymin_exp, xmin, ymin, xmax], axis=-1),
tf.stack([ymax, xmin, ymax_exp, xmax], axis=-1)
]
area_list = [
tf.cast(_get_box_area(b), tf.float32) for b in box_list
]
cumsum = imgproc.calc_cumsum_2d(score_map,
tf.concat(box_list, axis=1))
cumsum_list = [
cumsum[:, i * p:(i + 1) * p, :] for i in range(len(box_list))
]
# Compute the averaged cumsum inside each box.
cumsum_avg_list = [
tf.div(
cumsum_list[i],
tf.expand_dims(tf.maximum(_SMALL_NUMBER, area_list[i]),
axis=-1)) for i in range(len(box_list))
]
# The main box has to be valid, including the four shrinked boxes.
assert_op = tf.Assert(
tf.reduce_all(tf.greater(tf.stack(area_list, axis=-1), 0)),
["Check area of the main box failed:", area_list[0]])
with tf.control_dependencies([assert_op]):
grad_list = []
for i in [1, 2, 3, 4]:
grad = cumsum_avg_list[i] - cumsum_avg_list[i + 4]
grad_list.append(grad)
return purity_weight * cumsum_avg_list[0] + tf.reduce_min(
tf.stack(grad_list, axis=-1), axis=-1)
def _cumsum_avg(score_map, box):
ymin, xmin, ymax, xmax = tf.unstack(box, axis=-1)
area = tf.expand_dims((ymax - ymin) * (xmax - xmin), axis=-1)
return tf.div(imgproc.calc_cumsum_2d(score_map, box),
tf.maximum(_SMALL_NUMBER, tf.cast(area, tf.float32)))
