There were aspirations that the bottom up approach would work with enough data, but as I learned about the kind of long tail cases that we solved with radar/camera fusion, camera-only seemed categorically less safe.
easy edge case: A self driving system cannot be inoperable due to sunlight or fog.
a more hackernew worthy consideration: calculate the angular pixel resolution required to accurately range and classify an object 100 meters away. (roughly the distance needed to safely stop if you're traveling 80mph) Now add a second camera for stereo and calculate the camera-to-camera extrinsic sensitivity you'd need to stay within to keep error sufficiently low in all temperature/road condition scenarios.
The answer is: screw that, I should just add a long range radar.
there are just so many considerations that show you need a multi-modality solution, and using human biology as a what-about-ism, doesn't translate to currently available technology.