I’ve noted that Maryland state highway signs are being misidentified as speed limits. Some examples…
@asturksever could please you advise on what can (must?) be done here? If you know, of course.
I’m afraid I have no expertise in image recognition/AI. It would seem that the AI needs to be trained to recognize the differences between a typical speed limit sign and a Maryland state highway sign. How that’s done - IDK.
Clearly there are distinctions. The word “Maryland” on the state highway signs is one and the fact that Maryland state highway signs are accompanied by a rectangular direction sign above (i.e. North, South, East, West).
Sorry I couldn’t be more help. But I thought I would throw this out there in case anybody was interested.
A lot of traffic signs are misidentified. I would say as much as 10%+ are wrong. Sometimes it’s completely wrong, sometimes it’s close enough to a similar sign, sometimes it’s a non-standard sign, sometimes it’s not a sign at all. Signs with text (like a lot in US) are the worst, since the “AI” has no idea what they are. It’s much better in places like Europe where most signs are a standard pictogram. Unfortunately, there is no way to “flag” these as wrong.
At least it’s not as bad as object recognition. I stopped using it to find unmapped features when seemingly half the results identify dogs as benches. For example, I tried looking for fire hydrants, and I think something like 80% were misidentified.
In any case, what my point is is that the data is far, far from perfect and these examples are actually rather mild in comparison to some stuff.
As for what can be done - let us flag signs and objects as correct or wrong and use that real-world data to train the recognition algorithm.