Tesla put two visible light cameras in their cars, but didn't configure them to parallax and instead rely on radar that has noise problems and ultrasound which only works at close range. It was a mistake that has had consequences.
While I agree that it would be better to be using a multi-camera system for depth estimation, and there's no reason not to (because cameras are cheap and easy to build into bodywork of a vehicle), it shouldn't necessarily be needed.
Otherwise, you'd have one-eye sighted people, who drive, running into the same obstacles, right? But they don't - so how do they do it?
Probably motion cues, depth-from-motion, etc - something the brain does. From a cost-savings perspective, eliminating that extra camera will likely be looked at being done at some point, because over millions of cars manufactured, not installing that extra camera and wiring will save a ton of money and increase profitability.
But Tesla isn't quite there yet, and probably shouldn't eliminate it...
Furthermore, I'm not sure even LIDAR should be eliminated; the more sensors and sensor types, the better, ultimately. The problem is finding a LIDAR unit that has enough resolution and can be fitted to a vehicle (and doesn't look like a giant spinning top). Flash LIDAR units will probably be the solution, but they are still fairly expensive, and I don't know if they have the resolution needed yet, either.
Tesla's system has massively more users than all the other systems. They should get some benefit out of that data. I do worry that telsa will fail to make the software work well enough and get sued by all their existing customers (I'm a customer, but I want it to work :-)).
Most of the self-driving cars you see in bay area with lidars on top actually use those lidars to produce labeled data for vision-based distance estimator networks
His aversion to Lidar seems to have more to do with the cost implications on a vehicle that doesn't have a lot of room to raise the price or cut into the margins.