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)
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)