def testTransformBBoxes3DConsistentWithPoints(self):
num_boxes, num_points = 20, 100
points = tf.random_uniform((num_points, 3))
bboxes_3d = tf.random_uniform((num_boxes, 7))
in_bboxes = geometry.IsWithinBBox3D(points, bboxes_3d)
transforms = self._MakeTransformTestTranslationMatrices(1)[0]
points_transformed = geometry.TransformPoints(points, transforms)
bboxes_3d_transformed = geometry.TransformBBoxes3D(bboxes_3d, transforms)
in_bboxes_transformed = geometry.IsWithinBBox3D(points_transformed,
bboxes_3d_transformed)
with self.session() as sess:
actual_in_bboxes, actual_in_bboxes_transformed = sess.run(
(in_bboxes, in_bboxes_transformed))
self.assertAllEqual(actual_in_bboxes, actual_in_bboxes_transformed)
def add_labels(self, feature, labels, points_xyz):
"""Add 3d bounding box labels into the output feature map.
Args:
feature: A tf.Example feature map.
labels: A repeated car.open_dataset.Label proto.
points_xyz: A numpy array of shape [-1, 3] with the pointcloud. This is
used to calculate the number of points in each 3D bounding box.
"""
label_classes = []
label_ids = []
detection_difficulty_levels = []
tracking_difficulty_levels = []
bboxes = []
label_md = []
for label in labels:
box = label.box
bbox_3d = [
box.center_x, box.center_y, box.center_z, box.length,
box.width, box.height, box.heading
]
md = [
label.metadata.speed_x, label.metadata.speed_y,
label.metadata.accel_x, label.metadata.accel_y
]
label_md += md
bboxes += bbox_3d
label_classes += [label.type]
label_ids += [tf.compat.as_bytes(label.id)]
detection_difficulty_levels += [label.detection_difficulty_level]
tracking_difficulty_levels += [label.tracking_difficulty_level]
# Calculate the number of points in each ground truth box which are needed
# to fill in difficulty levels for each ground truth and to filter boxes
# with less points than a configurable minimum.
points_xyz = tf.convert_to_tensor(points_xyz, dtype=tf.float32)
bboxes_3d = tf.convert_to_tensor(np.array(bboxes).reshape(-1, 7),
dtype=tf.float32)
points_in_bboxes_mask = geometry.IsWithinBBox3D(points_xyz, bboxes_3d)
bboxes_3d_num_points = tf.reduce_sum(tf.cast(points_in_bboxes_mask,
tf.int32),
axis=0,
keepdims=False)
bboxes_3d_num_points = bboxes_3d_num_points.numpy().reshape([-1])
bboxes = np.array(bboxes).reshape(-1)
label_md = np.array(label_md).reshape(-1)
feature['labels'].int64_list.value[:] = label_classes
feature['label_ids'].bytes_list.value[:] = label_ids
feature['detection_difficulties'].int64_list.value[:] = (
detection_difficulty_levels)
feature['tracking_difficulties'].int64_list.value[:] = (
tracking_difficulty_levels)
feature['bboxes_3d'].float_list.value[:] = list(bboxes)
feature['label_metadata'].float_list.value[:] = list(label_md)
feature['bboxes_3d_num_points'].int64_list.value[:] = list(
bboxes_3d_num_points)
def _GetFilteredBoundingBoxData(kitti_data):
"""Given a single batch element of data, process it for writing.
Args:
kitti_data: A NestedMap of KITTI input generator returned data with a batch
size of 1.
Returns:
A NestedMap of all the output data we need to write per bounding box
cropped pointclouds.
"""
points = kitti_data.lasers.points_xyz
points_feature = kitti_data.lasers.points_feature
bboxes_3d = kitti_data.labels.bboxes_3d
bboxes_3d_mask = kitti_data.labels.bboxes_3d_mask
bboxes_3d = tf.boolean_mask(bboxes_3d, bboxes_3d_mask)
if 'points_padding' in kitti_data.lasers:
points_validity_mask = tf.cast(kitti_data.lasers.points_padding - 1,
tf.bool)
points = tf.boolean_mask(points, points_validity_mask)
points_feature = tf.boolean_mask(points_feature, points_validity_mask)
points_in_bboxes_mask = geometry.IsWithinBBox3D(points, bboxes_3d)
output_map = py_utils.NestedMap()
# Points and features contain the whole pointcloud, which we will use
# per box boolean masks later in _ToTFExampleProto to subselect data per box.
output_map.points = points
output_map.points_feature = points_feature
output_map.points_in_bboxes_mask = points_in_bboxes_mask
output_map.source_id = kitti_data.labels.source_id
# Add additional data
output_keys = [
'bboxes_3d',
'labels',
'texts',
'occlusion',
'difficulties',
'truncation',
]
for key in output_keys:
output_map[key] = tf.boolean_mask(kitti_data.labels[key],
kitti_data.labels.bboxes_3d_mask)
return output_map
def testIsWithinBBox3D(self):
num_points, num_bboxes = 19, 4
# rotate the first box by pi / 2 so dim_x and dim_y are swapped.
# The last box is a cube rotated by 45 degrees.
bboxes = tf.constant([[1.0, 2.0, 3.0, 6.0, 0.4, 6.0, np.pi / 2],
[4.0, 5.0, 6.0, 7.0, 0.8, 7.0, 0.0],
[0.4, 0.3, 0.2, 0.1, 0.1, 0.2, 0.0],
[-10., -10., -10., 3., 3., 3., np.pi / 4]],
dtype=tf.float32)
points = tf.constant(
[
[1.0, 2.0, 3.0], # box 0 (centroid)
[0.8, 2.0, 3.0], # box 0 (below x)
[1.1, 2.0, 3.0], # box 0 (above x)
[1.3, 2.0, 3.0], # box 0 (too far x)
[0.7, 2.0, 3.0], # box 0 (too far x)
[4.0, 5.0, 6.0], # box 1 (centroid)
[4.0, 4.6, 6.0], # box 1 (below y)
[4.0, 5.4, 6.0], # box 1 (above y)
[4.0, 4.5, 6.0], # box 1 (too far y)
[4.0, 5.5, 6.0], # box 1 (too far y)
[0.4, 0.3, 0.2], # box 2 (centroid)
[0.4, 0.3, 0.1], # box 2 (below z)
[0.4, 0.3, 0.3], # box 2 (above z)
[0.4, 0.3, 0.0], # box 2 (too far z)
[0.4, 0.3, 0.4], # box 2 (too far z)
[5.0, 7.0, 8.0], # none
[1.0, 5.0, 3.6], # box0, box1
[-11.6, -10., -10.], # box3 (rotated corner point).
[-11.4, -11.4, -10.], # not in box3, would be if not rotated.
],
dtype=tf.float32)
expected_is_inside = np.array([
[True, False, False, False],
[True, False, False, False],
[True, False, False, False],
[False, False, False, False],
[False, False, False, False],
[False, True, False, False],
[False, True, False, False],
[False, True, False, False],
[False, False, False, False],
[False, False, False, False],
[False, False, True, False],
[False, False, True, False],
[False, False, True, False],
[False, False, False, False],
[False, False, False, False],
[False, False, False, False],
[True, True, False, False],
[False, False, False, True],
[False, False, False, False],
])
assert points.shape[0] == num_points
assert bboxes.shape[0] == num_bboxes
assert expected_is_inside.shape[0] == num_points
assert expected_is_inside.shape[1] == num_bboxes
with self.session() as sess:
is_inside = sess.run(geometry.IsWithinBBox3D(points, bboxes))
self.assertAllEqual([num_points, num_bboxes], is_inside.shape)
self.assertAllEqual(expected_is_inside, is_inside)
def testIsWithinBBox3D(self):
num_points, num_bboxes = 19, 4
# rotate the first box by pi / 2 so dim_x and dim_y are swapped.
# The last box is a cube rotated by 45 degrees.
bboxes = tf.constant([[1.0, 2.0, 3.0, 6.0, 0.4, 6.0, np.pi / 2],
[4.0, 5.0, 6.0, 7.0, 0.8, 7.0, 0.0],
[0.4, 0.3, 0.2, 0.1, 0.1, 0.2, 0.0],
[-10., -10., -10., 3., 3., 3., np.pi / 4]],
dtype=tf.float32)
points = tf.constant(
[
[1.0, 2.0, 3.0], # box 0 (centroid)
[0.8, 2.0, 3.0], # box 0 (below x)
[1.1, 2.0, 3.0], # box 0 (above x)
[1.3, 2.0, 3.0], # box 0 (too far x)
[0.7, 2.0, 3.0], # box 0 (too far x)
[4.0, 5.0, 6.0], # box 1 (centroid)
[4.0, 4.6, 6.0], # box 1 (below y)
[4.0, 5.4, 6.0], # box 1 (above y)
[4.0, 4.5, 6.0], # box 1 (too far y)
[4.0, 5.5, 6.0], # box 1 (too far y)
[0.4, 0.3, 0.2], # box 2 (centroid)
[0.4, 0.3, 0.1], # box 2 (below z)
[0.4, 0.3, 0.3], # box 2 (above z)
[0.4, 0.3, 0.0], # box 2 (too far z)
[0.4, 0.3, 0.4], # box 2 (too far z)
[5.0, 7.0, 8.0], # none
[1.0, 5.0, 3.6], # box0, box1
[-11.6, -10., -10.], # box3 (rotated corner point).
[-11.4, -11.4, -10.], # not in box3, would be if not rotated.
],
dtype=tf.float32)
expected_is_inside = np.array([
[True, False, False, False],
[True, False, False, False],
[True, False, False, False],
[False, False, False, False],
[False, False, False, False],
[False, True, False, False],
[False, True, False, False],
[False, True, False, False],
[False, False, False, False],
[False, False, False, False],
[False, False, True, False],
[False, False, True, False],
[False, False, True, False],
[False, False, False, False],
[False, False, False, False],
[False, False, False, False],
[True, True, False, False],
[False, False, False, True],
[False, False, False, False],
])
assert points.shape[0] == num_points
assert bboxes.shape[0] == num_bboxes
assert expected_is_inside.shape[0] == num_points
assert expected_is_inside.shape[1] == num_bboxes
with self.session():
is_inside = self.evaluate(geometry.IsWithinBBox3D(points, bboxes))
self.assertAllEqual([num_points, num_bboxes], is_inside.shape)
self.assertAllEqual(expected_is_inside, is_inside)
# Add a batch dimension to the data and see that it still works
# as expected.
batch_size = 3
points = tf.tile(points[tf.newaxis, ...], [batch_size, 1, 1])
bboxes = tf.tile(bboxes[tf.newaxis, ...], [batch_size, 1, 1])
with self.session():
is_inside = self.evaluate(geometry.IsWithinBBox3D(points, bboxes))
self.assertAllEqual([batch_size, num_points, num_bboxes],
is_inside.shape)
for batch_idx in range(batch_size):
self.assertAllEqual(expected_is_inside, is_inside[batch_idx])
