Benedict Evans, a general partner at Andreessen Horowitz, claims that “almost all autonomy” projects are using lidar for SLAM, and that not all of them use HD maps.
An MIT group is testing self-driving cars on public roads without HD maps.
My question is whether the difference in error between lidar and cameras is significant. Benedict Evans and others claim that lidar is necessary for accurate enough SLAM in self-driving cars, but at first glance the KITTI benchmark data seems to contradict that claim. I want to confirm or refute that impression.
The KITTI Vision benchmark leaderboard for visual odometry/SLAM methods shows a lidar-based method called V-LOAM in first place, and a stereo camera-based method called SOFT2 in fourth place. They have the same rotation error, and a percentage point difference of 0.05 in their respective translation errors.
Is a 0.05 percentage point difference in translation accuracy large or insignificant, when it comes to an autonomous car navigation?
The KITTI Vision benchmark leaderboard for odometry/SLAM methods: