In the Netherlands we have three types op traffic signs for cycleways. 2 of them are recognised but the one that is usually the most scenic is not. It is very often blurred . It is sign G13 with the word “FIETSPAD” on it.
See this list of NL traffic signs
See this photo on which it is visible and the next on which it is blurred.
A general answer is yes, it is possible to add these signs to be recognised. And we will extend the algorithm in the next iterations, but it will take some time. That is because it’s not only a case of adding the sign to the list, but also training the algorithm to recognise it in different conditions. Need examples for that so everyone, please feel free to capture these whenever you can!
In the example given, there isn’t any blurring anymore, that’s a first step
I can tag “Bicycle” and “Bicycle Lane”… is there a list somewhere what is expected for which word? Or in other words a list of all tags and what is expected to be tagged?
Ack - these are good examples of where today’s CV detections fall short. In the longer run we would like to improve the models to better handle these cases. In the short term the models have been static for a bit, so this is something we’ll need to prioritize for.
Is there any specific reason why parking traffic sign (regulatory‑‑no‑parking‑‑g* and perhaps others too) detection has been disabled or any such detections are hidden from API queries for years? I wish these would come back.
Why aren’t existing sequences, that are touched, affected, or crossed (what ever you call it) by new sequences, also subject to the latest training dataset traffic sign detection?