However, humans get by driving reasonable safely using just vision. I'm totally a computer vision layman, but couldn't you correctly interpret this situation with binocular vision?
The advantage is mainly in software. For example, I have a spot on my daily commute where I need to yield, but thick vegetation blocks the line of sight. I have to look for the signs of a moving car through the bush where there are some clearings, then wait to give the cars that I can't see the time to appear on the other side, if they exist, and proceed only then. I seriously doubt an image recognition system can do this.
Radar would do an even better job. There's no reason why cars shouldn't have radars and drive better than humans, but yes, they should be able to solve that by vision alone.
Not really. A purely vision-based system probably needs to see this example in training labeled as "the back of a car" in order to differentiate it from reality.
A sensor-diverse system could detect the metal on the car and determine it is a car without ever having seen such an example during training.
And, this is just one edge case. The number of edge cases is infinite.
The point in the article is that a diverse sensor array is advantageous because it will cover more edge cases.
A purely vision based system will always have limitations compared to a system with more kinds of sensors.
Have you take a look at depth based feature recognition? It will automatically recognize everything as a car, and will tell the error of the 2D feature extraction every time.
It's also extremely CPU intensive, so people don't apply it as often. It's a clear example of AI being hold down due to limitations in computer power.
Getting 3d from 2d images is certainly interesting and will have its applications. It's less accurate and more computationally demanding than using 3d input directly, however. Cars can and should have more than 2d input.
The point of the article is that using a diverse array of sensors in cars will help the algorithm make better predictions.
There are many edge cases out there that aren't covered,and we shouldn't expect engineers to plan for them all. What we want to do is throw a lot of raw input at an algorithm and run it through a ton of simulations. All steps of this are non-trivial, but with more data, it is easier.
It's news when a self-driving car runs a red light, but pretty normal for humans.