3 ms·
What I don't get: if deep learning can recognize a crosswalk with 99.99% accuracy, that's very impressive. Yet it is not enough.
by amelius 5y ago
What I don't get: if deep learning can recognize a crosswalk with 99.99% accuracy, that's very impressive. Yet it is not enough.
- vanviegen 5y agoDo human drivers spot crosswalks in time with higher accuracy? That seems very unlikely to me. Still, as long as the driver (human or otherwise) sees the pedestrian and vice versa, an accident can usually be avoided, with perhaps some angry shouting.
- bastawhiz 5y agoI can imagine this being at least two orders of magnitude higher than most human drivers (one in ten thousand misses versus one in one hundred). How many orders of magnitude better does it need to be before it's good enough?
- robonerd 5y agoMissing one in a hundred intersections? That's insane. Most human drivers are much better than that and wouldn't keep their license (or their life) for long if they weren't.
- eggsmediumrare 5y agoI dunno man, I live next to a school... And based on my observations, I think the actual occurence is probably closer to 1/100 than 1/10000.
- amelius 5y agoMy 99.99% was just a number for the sake of argument. The real accuracy of classification networks is usually in the range of 90-99%. It's not bad for a lot of tasks, but for self-driving I'd say it's far from sufficient.
- Dave_Rosenthal 5y agoYou get it correctly. Deep learning isn't inherently good for cases where you need to be right 100% of the time. But we don't have a better tool so the industry is trying to see how far they can push it. One big open question in self driving is: Do you have to be able to see/classify everything in the scene all the time to be a good driver? I think the answer is no (see: humans.) But, because most self driving stacks in the industry have chosen "3d classified representation of all of the objects in the scene" as a fundamental interface between perception and planning stages, there is a big challenge. (e.g. Representing uncertainty usefully is exceptionally difficult with this interface, among other things.)