This is why you should use radar and/or lidar, not just cameras...
That they were planning to remove radar was clear a while before but the electronic component shortage made it a higher priority problem.
Based on this video I’m not sure the radar would’ve had anything to catch. It seems like the plane hit the top of the car, not that the car hit the landing gear.
However, I do think radar should have caught it, because radar has almost like a cone like view in front. If the car was far enough, the cone would've spotted the height.
The camera performance is like the newer starlight sensors in terms of seeing in the dark, and they seem to process from raw data. Essentially, once they crack all the quirks, I am now a firm beleiver that it will better than the average joe with limited angulars vision+rotating head vs. surround stereoscopic vision.
Not saying I 100% agree with launching a beta program in the wild like this, but to their defense, given the current government driving paperwork and legislation to insane levels of unproductivity, innovators and challengers kinda have to play it out like that if they want to disrupt sectors (in the end) for the better.
Not saying the end justifies the means, but I can understand the logic.
The main issue with vision-only FSD is that visual artifacts don't convey high degrees of certainty about the solid nature of things.
If the AI doesn't recognize an object as being an object it can't add it to the list of things to worry about. If it's augmented with LIDAR and radar it doesn't need to recognize an airplane to know there's a solid object in its path.
It's dumb edge cases that cause problems. Vision-only FSD leaves a lot of edge cases on the table with the hope that AI magic will save the day.
I think maybe someone dazzled you with some excessive jargon. Extracting point clouds from stereoscopic cameras isn't a Tesla invention, even though they've invested a lot into trying to make theirs work. This is some computer vision 101 stuff, but nobody calls it "converting photons".
> Essentially, once they crack all the quirks, I am now a firm beleiver
Well obviously if they solve all of the problems then we'd all be believers too.
Kinect point cloud was a total mess, never worked with more than a few cameras and had a LOT of restriction and artifacts, what I meant by impressive, imagine a moving 6 camera kinect with clean signals, that doesn't fail on specific texture or reflection because it track in 4D + AI.
Probably the people downrating are the same thinking the Tesla FSD AI is a kinect contraption.
I just spent the time watching the vision segment of that video, and while the presenter uses the brain as a simile to motivate neural network architecture, it actually is not modeled at all on how primate visual cortex works. Seems some tweeps also got caught up in this hype. I think those few slides were really just to show the complexity of the problem.
For parties interested in actual models of biological vision, https://serre-lab.clps.brown.edu/wp-content/uploads/2014/10/... is a good resource.
The biggest problem I see is not if they will solve it (that's just a question of time), its how they communicate their optimism and "almost there" but always pushing further. That's more than annoying (at best) and contributing extensively to the negativity. I guess they are trying to get people excited, attract talent, etc.
There may well be valid arguments for radar/lidar superiority over cameras, but the line of reasoning you're suggesting isn't in that category.
I don't think anyone would argue that cameras couldn't do the same job if capable of realizing depth and re-assessing from different angles when confronted with ambiguity... but show me what a system like that would look like.
A human that is parking somewhere ambiguously often gets out of their car to inspect the area with their binocular vision from multiple angles; this is common among trailer owners.
The vision system on a vehicle is tasked with a harder problem in many ways -- it's trying to solve the ambiguity of the scene purely visually with mostly-stationary cameras and prior knowledge.
> realizing depth and re-assessing from different angles
A car with three or four widely spaced cameras can get way better visual information than a human can. (At least in the forward direction. Add more cameras for 360.) To the point that you could probably give the human a VR set tied into the cameras and have them do a better job because you can make the pillars transparent. (Might need a special VR set with extra-wide field of view.)
> but show me what a system like that would look like.
I'm guessing not so much more wacky as the lidar based cars I see driving through my neighborhood. The accident in the original link may well have been caused by an inadequate vision based system (I don't know), but just because an accident happened with a single vision implementation doesn't mean that somehow Lidar/Radar would necessarily be superior. That conclusion, tossed off by the original comment, is just wrong as it stands.
> A human that is parking somewhere ambiguously often gets out of their car to inspect the area with their binocular vision from multiple angles;
In most cases, but not always. I have an old friend that was completely blind in one eye: still legally licensed to drive and successfully navigated the environment better than a lidar/radar/artificial vision system still can. Even in bad weather conditions such as snow and rain (and night). Will that hold true over time? Hard to say, but I'd bet right now that the larger problem is processing the received information regardless of how the information was acquired.
The problem with the original statement is that it reminded me of the line of reasoning that goes, "there are aspects of the world that science has no answer for, therefore God." or "there was an abnormally cold day in spring, therefore no global warming" Tesla hasn't implemented a sufficiently robust visual driving system to avoid running into parked airplanes, therefore Lidar/Radar. It simply doesn't follow. Much more is needed.
Ultimately, to date, the only sufficient input for fully independent driving that we have is vision based. Lidar/Radar may ultimately be superior, but a vision based system is so far the only generally accepted method for driving... even if it is biological vision rather than computer vision.