Tesla went nothing-but-nets (making fusion easy) and Chinese LIDAR became cheap around 2023, but monocular depth estimation was spectacularly good by 2021. By the time unit cost and integration effort came down, LIDAR had very little to offer a vision stack that no longer struggled to perceive the 3D world around it.

Also, integration effort went down but it never disappeared. Meanwhile, opportunity cost skyrocketed when vision started working. Which layers would you carve resources away from to make room? How far back would you be willing to send the training + validation schedule to accommodate the change? If you saw your vision-only stack take off and blow past human performance on the march of 9s, would you land the plane just because red paint became available and you wanted to paint it red?

I wouldn't completely discount ego either, but IMO there's more ego in the "LIDAR is necessary" case than the "LIDAR isn't necessary" at this point. FWIW, I used to be an outspoken LIDAR-head before 2021 when monocular depth estimation became a solved problem. It was funny watching everyone around me convert in the opposite direction at around the same time, probably driven by politics. I get it, I hate Elon's politics too, I just try very hard to keep his shitty behavior from influencing my opinions on machine learning.

depth estimation is but one part of the problem— atmospheric and other conditions which blind optical visible spectrum sensors, lack of ambient (sunlight) and more. lidar simply outperforms (performs at all?) in these conditions. and provides hardware back distance maps, not software calculated estimation

Lidar fails worse than cameras in nearly all those conditions. There are plenty of videos of Tesla's vision-only approach seeing obstacles far before a human possibly could in all those conditions on real customer cars. Many are on the old hardware with far worse cameras

Interesting, got any links? Sounds completely unbelievable, eyes are far superior to the shitty cameras Tesla has on their cars.

Yea.. not holding my breath on links to superman tesla cameras performing better than eyes

There's a misconception that what people see and what the camera sees is similar. Not true at all. One day when it's raining or foggy, have some record the driving, through the windshield. You'll be very surprised. Even what the camera displays on the screen isn't what it's actually "seeing".

Always thought the case was for sensor redundancy and data variety - the stuff that throws off monocular depth estimation might not throw off a lidar or radar.

Monocular depth estimation can be fooled by adversarial images, or just scenes outside of its distribution. It's a validation nightmare and a joke for high reliability.

It isn't monocular though. A Tesla has 2 front-facing cameras, narrow and wide-angle. Beyond that, it is only neural nets at this point, so depth estimation isn't directly used; it is likely part of the neural net, but only the useful distilled elements.

I never said it was. I was using it as a lower bound for what was possible.

> but monocular depth estimation was spectacularly good by 2021

It's still rather weak and true monocular depth estimation really wasn't spectacularly anything in 2021. It's fundamentally ill posed and any priors you use to get around that will come to bite you in the long tail of things some driver will encounter on the road.

The way it got good is by using camera overlap in space and over time while in motion to figure out metric depth over the entire image. Which is, humorously enough, sensor fusion.

It was spectacularly good before 2021, 2021 is just when I noticed that it had become spectacularly good. 7.5 billion miles later, this appears to have been the correct call.

What are the techniques (and the papers thereof) that you consider to be spectacularly good before 2021 for depth estimation, monocular or not?

I do some tangent work from this field for applications in robotics, and I would consider (metric) depth estimation (and 3D reconstruction) starting to be solved only by 2025 thanks to a few select labs.

Car vision has some domain specificity (high similarity images from adjacent timestamps, relatively simpler priors, etc) that helps, indeed.

It doesn't solve the "Coyote paints tunnel on rock" problem though.

IIRC, that was only ever a problem for the coyote, though.

Source: not a computer vision engineer, but a childhood consumer of looney toons cartoons.

Time for a car company to call itself "ACME" and the first model the "Road Runner".