When evaluating performance on KITTI depth data, are you using raw sparse lidar measurements, or post-processed dense depth map as ground truth?