Humans don’t regress per-pixel depth or use convolutions and region proposals to draw bounding boxes around objects. They don’t function on models with fixed weights trained by backprop either. The idea that “vision only” somehow more closely resembles how humans drive quickly falls apart if you inspect the internals of these systems.
The similarity is in the problem to be solved, not the details of the compute pipeline. Depth must be inferred somehow, rather than measured by actively interacting with the surface, as it is in LiDAR.
If you really wanted to make this argument, you wouldn’t even want to bother inferring depth, since that’s not what humans do, not directly at least. If you’re actually trying to obtain a depth map as part of your pipeline, LIDAR (or LiDAR + vision feeding into a denser depth prediction model) would always be a better strategy, cost aside, since determining depth from images is an ill posed problem.
My claim is that humans use their eyes as a primary input for driving. I don't think it's controversial. We don't let eyeless people drive. Eyes do not shoot out lasers.
I think the comparison comes from the fact that humans infer a 3D map from stereo vision whereas LiDAR to some extent gives you the ground truth 3D map. You’re right that it falls apart pretty quickly though.
Except Tesla isn’t inferring a 3D map from stereo vision either, at least not outside of the front facing cameras - they’re using monocular depth prediction.
Neither are humans. Our eyes are so close together there's almost no disparity between the eye images beyond a handful of meters. We do 3D by inference beyond the near-field.
