Hacker News new | ask | show | jobs
by balia 1883 days ago
This seems it will be applied to self driving technology in a huge way.

All of these companies doing advanced AI vision detection, classification etc... they’re really hard problems. But the whole challenge will eventually become nullified when every single road sign, landmark, and the road itself are tagged internally with metadata.

Instead of trying to decipher how does this PNG of a speed limit sign translate into a number, the number metadata would be encoded in the sign.

Looking at it this way, having advanced vision AI tech is only a competitive advantage in the short term

6 comments

My car already does a pretty good job of showing the speed limits (and it's definitely based on computer vision/ works offline and in places where map metadata is poor). And it doesn't have "self driving", just a limited "pilot assist".

Not sure why reading road signs is used as and example of "extremely hard thing to do" - there are other way harder things that cars can't currently do. E.g. figuring out the right speed to negotiate a curve using computer vision alone (without relying on detailed maps/gps, ie "metadata").

Actually I think road speeds for various curves are pretty standardized, so it shouldn't be that hard for the car to estimate the correct speed based on the curve.
It's a tad more complicated than that - it needs to consider also driving conditions, any hazards on the road. Probably can't be done on computer vision alone, you need at least some sensors to get a "feel" of how good is the road grip.
My Audi tries to account for curves in the road when running on cruise control through the onboard nav. It’s horrendously conservative in its opinion of the maximum safe speed to the point I find it dangerous to rely upon. If I left it to its own devices, other cars would be passing me as if I were standing still. Otherwise the driver assist features are fairly slick.
How does your car handle snow-covered road signs?
The difficulty is getting the kid playing in the road to correctly encode their metadata; not so much the posted signs.
You see, that's what the chips in the vaccines are for /s
> But the whole challenge will eventually become nullified when every single road sign, landmark, and the road itself are tagged internally with metadata

Including every pothole, pedestrian, stray deer, abandoned shopping cart and fallen tree branch?

What about collaboratively tagging drivers with metadata - I suspect that could be pretty useful... ;-)
> every single road sign, landmark, and the road itself are tagged internally with metadata.

Who's going to pay for that? The public? In order to provide returns to private shareholders?

I wonder if it would be cheaper for all these companies to go and pay for a sticker/stamp of some kind put on all the signs in a city. Figure out a standard and put them everywhere.

This would give them time to get the AI vision detection algorithms figured out.

what's the point of putting it on the signs

when you can create virtual sign map that every city/road maintenance HAS to update?

Personally, I'd be more comfortable by making the signs themselves easier for a machine to read rather than to build/maintain what is basically a form of dead reckoning.
>when you can create virtual sign map that every city/road maintenance HAS to update?

Who can do that?

society should, just like roads.
At least in the US, "society" doesn't build and maintain roads. Thousands of individual city, county and state governments do.
ain't they just society representatives?

anyway meant the same

> the number metadata would be encoded in the sign.

Oh yea that's not going to backfire at all is it? Lemme just work out how to frig it to state the max speed limit is zero and create an automated car pile-up.