def ExtractNpContent(np_dict, calib):
"""Parse saved np arrays and convert 3D bboxes to camera0 coordinates.
Args:
np_dict: a dict of numpy arrays.
calib: a parsed calibration dictionary.
Returns:
A tuple of 6 ndarrays:
- location_camera: [N, 3]. [x, y, z] in camera0 coordinate.
- dimension_camera: [N, 3]. The [height, width, length] of objects.
- phi_camera: [N]. Rotation around y-axis in camera0 coodinate.
- bboxes_2d: [N, 4]. The corresponding 2D bboxes in the image coordinate.
- scores: [N]. Confidence scores for each box for the assigned class.
- class_ids: [N]. The class id assigned to each box.
"""
bboxes = np_dict["bboxes"]
scores = np_dict["scores"]
class_ids = np_dict["class_ids"]
bboxes_2d = np_dict["bboxes_2d"]
# Transform from velodyne coordinates to camera coordinates.
velo_to_cam_transform = kitti_data.VeloToCameraTransformation(calib)
location_cam = np.zeros((len(bboxes), 3))
dimension_cam = np.zeros((len(bboxes), 3))
rotation_cam = np.zeros((len(bboxes), 1))
for idx, bbox in enumerate(bboxes):
location_cam[idx, :], dimension_cam[idx, :], rotation_cam[idx, :] = (
kitti_data.BBox3DToKITTIObject(bbox, velo_to_cam_transform))
return location_cam, dimension_cam, rotation_cam, bboxes_2d, scores, class_ids
def testKITTIObjToBBoxAndInverse(self):
objects = kitti_data.LoadLabelFile(self._label_file)
calib = kitti_data.LoadCalibrationFile(self._calib_file)
for obj in objects:
bbox3d = kitti_data._KITTIObjectToBBox3D(
obj, kitti_data.CameraToVeloTransformation(calib))
location, dimensions, rotation_y = kitti_data.BBox3DToKITTIObject(
bbox3d, kitti_data.VeloToCameraTransformation(calib))
self.assertAllClose(obj['location'], location)
self.assertAllClose(obj['dimensions'], dimensions)
self.assertAllClose(obj['rotation_y'], rotation_y)