The paper did not compare cameras to lidar. However, we can use the numbers from the paper to compare cameras to lidar. The authors found that at low driving speeds, a four-camera system can achieve a localization accuracy of under 10 cm (3.9 in), just over the length of a credit card. Lidar has an accuracy of 1.5 cm (0.6), about the width of a finger.
Based on these numbers, my hunch is that lidar is overkill and cameras are good enough — crucially, as long as cameras are as accurate at high speeds. (That’s something I’m still trying to figure out.) If the car leaves a margin of error of 45 cm (about 18 inches) and it miscalculates distance by 20 cm, it will still be 25 centimetres (9.8 inches) away from the object.
Further evidence to support my hunch is that
one study found human drivers are only able to park with about 10 cm of lateral accuracy (left/right), and much less longitudinal accuracy (forwards/backwards).
With parking, you generally want to be closer than 45 cm to the curb, so the margin of error should be less. However, it isn’t a big deal if the car touches the curb, and in general auto-parking seems like it’s almost already a solved problem. In parking and other low speed situations, a HW2 Tesla can use the ultrasonics as well as the cameras.
Finally, Tesla uses eight cameras, whereas the experiments used only four cameras. Tesla augments its multi-camera system with GPS, unlike the system in the paper. From what I understand, the localization software used in the four-camera system consisted of hand-coded algorithms. It’s possible that deep neural networks could get better performance. At low speeds, Tesla can check its cameras against ultrasonics and radar, and at high speeds can check its cameras against radar.
So, with better hardware and possibly better software, as well as more time, more money, and more engineering talent, it’s likely that Tesla can achieve even better accuracy than was achieved in the experiments. That would close the gap even further between under 10 cm for cameras and 1.5 cm for lidar. The only remaining question is how much this level of accuracy deteriorates at high driving speeds. That’s why I’m currently looking into Tesla’s camera hardware and the software Tesla could be using to correct visual artifacts that occur at high speeds.