def testVeloToImagePlaneTransformation(self):
objects = kitti_data.LoadLabelFile(self._label_file)
calib = kitti_data.LoadCalibrationFile(self._calib_file)
# Only apply to object 0.
obj = objects[0]
bbox3d = kitti_data._KITTIObjectToBBox3D(
obj, kitti_data.CameraToVeloTransformation(calib))
# Convert to corners in our canonical space.
corners = geometry.BBoxCorners(
tf.constant([[bbox3d]], dtype=tf.float32))
with self.session():
corners_np = self.evaluate(corners)
corners_np = corners_np.reshape([8, 3])
# Add homogenous coordinates.
corners_np = np.concatenate([corners_np, np.ones((8, 1))], axis=-1)
# Apply the velo to image plane transformation.
velo_to_img = kitti_data.VeloToImagePlaneTransformation(calib)
corners_np = np.dot(corners_np, velo_to_img.T)
# Divide by the last coordinate to recover pixel locations.
corners_np[:, 0] /= corners_np[:, 2]
corners_np[:, 1] /= corners_np[:, 2]
# Obtain 2D bbox.
min_x = np.min(corners_np[:, 0])
max_x = np.max(corners_np[:, 0])
min_y = np.min(corners_np[:, 1])
max_y = np.max(corners_np[:, 1])
bbox = [min_x, min_y, max_x, max_y] # left, top, right, bottom.
# This should correspond to the GT bbox in obj['bbox'].
# We use atol=0.1 here since they should close to the nearest pixel.
self.assertAllClose(bbox, obj['bbox'], atol=0.1)
def _ReadObjectDataset(root_dir, frame_names):
"""Reads and parses KITTI dataset files into a list of TFExample protos."""
examples = []
total_frames = len(frame_names)
for frame_index, frame_name in enumerate(frame_names):
image_file_path = os.path.join(root_dir, 'image_2',
frame_name + '.png')
calib_file_path = os.path.join(root_dir, 'calib', frame_name + '.txt')
velo_file_path = os.path.join(root_dir, 'velodyne',
frame_name + '.bin')
label_file_path = os.path.join(root_dir, 'label_2',
frame_name + '.txt')
example = tf.train.Example()
feature = example.features.feature
# frame information
feature['image/source_id'].bytes_list.value[:] = [frame_name]
# 2D image data
encoded_image = tf.gfile.Open(image_file_path).read()
feature['image/encoded'].bytes_list.value[:] = [encoded_image]
image = np.array(Image.open(io.BytesIO(encoded_image)))
assert image.ndim == 3
assert image.shape[2] == 3
image_width = image.shape[1]
image_height = image.shape[0]
feature['image/width'].int64_list.value[:] = [image_width]
feature['image/height'].int64_list.value[:] = [image_height]
feature['image/format'].bytes_list.value[:] = ['PNG']
# 3D velodyne point data
velo_dict = kitti_data.LoadVeloBinFile(velo_file_path)
point_list = velo_dict['xyz'].ravel().tolist()
feature['pointcloud/xyz'].float_list.value[:] = point_list
reflectance_list = velo_dict['reflectance'].ravel().tolist()
feature[
'pointcloud/reflectance'].float_list.value[:] = reflectance_list
# Object data
calib_dict = kitti_data.LoadCalibrationFile(calib_file_path)
if tf.gfile.Exists(label_file_path):
# Load object labels for training data
object_dicts = kitti_data.LoadLabelFile(label_file_path)
object_dicts = kitti_data.AnnotateKITTIObjectsWithBBox3D(
object_dicts, calib_dict)
else:
# No object labels for test data
object_dicts = {}
num_objects = len(object_dicts)
xmins = [None] * num_objects
xmaxs = [None] * num_objects
ymins = [None] * num_objects
ymaxs = [None] * num_objects
labels = [None] * num_objects
has_3d_infos = [None] * num_objects
# 3D info
occlusions = [None] * num_objects
truncations = [None] * num_objects
xyzs = [None] * num_objects
dim_xyzs = [None] * num_objects
phis = [None] * num_objects
for object_index, object_dict in enumerate(object_dicts):
xmins[object_index] = object_dict['bbox'][0]
xmaxs[object_index] = object_dict['bbox'][2]
ymins[object_index] = object_dict['bbox'][1]
ymaxs[object_index] = object_dict['bbox'][3]
labels[object_index] = object_dict['type']
has_3d_infos[object_index] = 1 if object_dict['has_3d_info'] else 0
occlusions[object_index] = object_dict['occluded']
truncations[object_index] = object_dict['truncated']
xyzs[object_index] = object_dict['bbox3d'][:3]
dim_xyzs[object_index] = object_dict['bbox3d'][3:6]
phis[object_index] = object_dict['bbox3d'][6]
feature['object/image/bbox/xmin'].float_list.value[:] = xmins
feature['object/image/bbox/xmax'].float_list.value[:] = xmaxs
feature['object/image/bbox/ymin'].float_list.value[:] = ymins
feature['object/image/bbox/ymax'].float_list.value[:] = ymaxs
feature['object/label'].bytes_list.value[:] = labels
feature['object/has_3d_info'].int64_list.value[:] = has_3d_infos
feature['object/occlusion'].int64_list.value[:] = occlusions
feature['object/truncation'].float_list.value[:] = truncations
xyzs = np.array(xyzs).ravel().tolist()
feature['object/velo/bbox/xyz'].float_list.value[:] = xyzs
dim_xyzs = np.array(dim_xyzs).ravel().tolist()
feature['object/velo/bbox/dim_xyz'].float_list.value[:] = dim_xyzs
feature['object/velo/bbox/phi'].float_list.value[:] = phis
# Transformation matrices
velo_to_image_plane = kitti_data.VeloToImagePlaneTransformation(
calib_dict)
feature['transform/velo_to_image_plane'].float_list.value[:] = (
velo_to_image_plane.ravel().tolist())
velo_to_camera = kitti_data.VeloToCameraTransformation(calib_dict)
feature['transform/velo_to_camera'].float_list.value[:] = (
velo_to_camera.ravel().tolist())
cam_to_velo = kitti_data.CameraToVeloTransformation(calib_dict)
feature['transform/camera_to_velo'].float_list.value[:] = (
cam_to_velo.ravel().tolist())
examples.append(example)
if frame_index % 100 == 0:
logging.info('Processed frame %d of %d.', frame_index,
total_frames)
return examples