Hacker News Reader

Comment by jeffreygoesto | Hacker News Reader

Tesla only uses Cameras (trained for "normal appearance" of cars, that is front and back views) and Radar (notoriously high false positive rate for standstill targets aka clutter, they need to be suppressed to be driving without hard braking every now and then), no Lidar. This overturned truck would have been difficult to incorporate into a training set for semantic labeling if you want to have a safe reaction for all kinds of unusual objects together with a reasonable false positive rate. Similar classification failures in all those cases. Can't be solved reliably with just vision and Radar.

View on HN

Can't radar and cameras augment each other and help to solve false positives? I just don't understand, it's pretty simple thing to notice, humans can solve it with vision only. It doesn't need to label what it is exactly, it just needs to understand that there's something, and it's probably not a good idea to plow in it full speed. I'm very positive about what Tesla is doing overall but this issue seems to be unresolved for years, and the way they handle these cases information-wise is also bad.

Computer vision is far behind human vision. Human vision includes the human brain, which we have not been able to come close to replicating. I am working on a years long computer vision project right now. We still struggle to do basic tasks that humans can do easily.

Making a computer vision system for a car that always knows what images mean the road is clear and what images mean the road is blocked with 100% accuracy is very hard.

That said, Tesla has also said they can determine the 3D geometry of the scene using only computer vision, which is certainly a known task (called photogrammetry or structure from motion). However it seems their algorithm failed in this case to accurately judge depth. With a clear picture and a well understood camera system it is pretty do-able, but I haven't looked at this article so perhaps the footage was not sufficiently clear.

> it's pretty simple thing to notice, humans can solve it with vision only

Yes.

Unfortunately, this wasn't human vision: it was computer vision.

When both systems succeed, they by definition succeed similarly: I say by definition, because human vision is the only judge of computer vision's success.

When they fail, they fail completely differently.

You can read the article, and see the video yourself, and the answer is no: Tesla couldn't manage reliable detection of "big object blocking the way".

yeah, but Tesla being unable to do it doesn't mean that someone can't.

It's a hard problem, but a system with more sensory inputs than the human optical system shouldn't solve the same problem worse -- this points to a real issue in the problem solving methodology, but doesn't do much to condemn the idea of camera/radar fusion.

Now, if the question was "Can Tesla do this reliably?" -- different answer.

This is what symbolic AI skeptics in the 90's dismissed as 'brittle' solutions. You engineer an AI system, then find it fails under some unforeseen situation. You fix for that situation and find it fails under some other unforeseen situation.

I was driving on the highway the other day and there was a large black object ahead in my lane. I was about to take evasive action when the object shifted slightly. The motion was enough for my brain to sense it was a lightweight object (not a human fallen on the road) and a further movement in a puff of wind to establish it was an empty garbage bag rolling on the highway. No need for evasive action.

On the other hand if I had seen other debris/leaves tumbling in the wind but the black object was stationary my brain would have sensed it as a potentially heavy object and taken evasive action.

I remain skeptical that we can if-then-else engineer our way to self-driving until cars have the common sense adaptive behavior that humans have.

The human system includes the human brain, which is an important part. We can easily make optical sensors that have better visual acuity than the human eye, but "scene understanding" continues to be a task where humans dramatically beat computers. The brain is an important part of this.

There's no question we'll get there eventually. But what's the timeline?

The first principles argument can be made just as well about setting up two cameras and some robotic hands to converse on HN and acquire karma.

It's perfectly reasonable for a system with more sensory inputs than a human to fail at a task a human does with just the one.

Do you think the car understands cause and effect? Do you think the car actually groks object persistence? Do you think that the car has any sense of value attached to it's own existence? Or an instinct to increase the amount of time it has to resolve a reading it doesn't understand?

This is the failing of your lab trained neural network. It is only trained and proficient in what it knows, and only that. It doesn't have a sense of self. It doesn't have intuition, instinct for survival, or any of the millions of other simultaneous neural subprocesses that even the simplest biological creatures possess to keep themselves alive. It can't forget, neither can it learn in it's runtime state. It's a crapshoot whether it'll do everything it is supposed to how it is supposed to just like it is when a human (coincidently, another neural network driven creature) gets behind the wheel.

You won't build a perfect neural network for X just like you can't make a perfect human by selective lobotomy. There is only so much neural ductility that can be accommodated.

I read the article and saw the video. I'm answering to the comment talking about the general problems of the approach.

Humans do have range-finding capabilities due to binocular vision.

Tesla doesn’t always suppress radar. My Tesla will occasionally phantom brake due to long sweeping left hand curves in the carpool lane with the adjacent lane stopped or low hanging road signs.

> Radar (notoriously high false positive rate for standstill targets aka clutter, they need to be suppressed to be driving without hard braking every now and then)

Really? My adaptive cruise on my bog-standard GM Volt seems to have no problem seeing static obstacles with it's "primitive" radar.

The problem is not "seeing" the stationary obstacles, the problem is distinguishing ones in the lane from those outside. Issues arise when you try to avoid false positives (such as cars in a different lane in a turn, etc.).

> the problem is distinguishing ones in the lane from those outside.

My vehicle is doing adaptive cruise to the car in front of it--including on banked curves and in Los Angeles traffic--sometimes dense-sometimes empty. This means that it is working out which car is in my lane--probably via beam steering and corresponding to my steering wheel position.

This doesn't seem to be an impossible problem. This seems to be a Tesla problem.

This article has a longer discussion on the limitations of radar for emergency braking systems https://arstechnica.com/cars/2018/06/why-emergency-braking-s...

This one discusses that the Volvo system has the same limitation, and implies that most others do too https://www.wired.com/story/tesla-autopilot-why-crash-radar/

The one in mine signals collision warnings on metal signs above the road. Although in fairness it's never triggered emergency braking.

Reply on news.ycombinator.com