All the computer vision work we did so far around the road signs has been focused on France but Panoramax is designed to extend everywhere.
Last friday, I got in touch with the netherlands and we talked about this as we have been adding semantic tags to previously detected road signs on a large number of recently uploaded pictures.
The tag uses the french references, like osm|traffic_sign=FR:B1 for a “no entry” sign.
Another limit is that the classification model has been trained to classify the official french signs as some foreign signs will not be classified correctly :
So I’ve done the following :
- extract road signs only located in the Netherlands
- get a list of sequences whose geometry is fully in the country
- download low-resolution pictures
- use the usual detect_face_plate_sign on them to detect signs and extract cropped versions
- automatic pre-clustering based on visual similarity (this helps the manual process but cannot replace it)
I got more than 23000 signs extracted, and published them on Panoramax/unclassified_nl_road_signs · Datasets at Hugging Face
Then…
- manual annotation of around 10.000 samples
- create one folder for eash type of sign (I’ve used generic names but they can be updated to official ref later by just renaming the folders)
- put in each folder enough examples (it is amazing to see that it starts to work pretty well with as little as 10 samples)
- training by fine tuning the classify_fr_road_signs model with the manualy annotated pictures (the idea is to take advantage of previous trainings on french road signs, the model already “knows” pretty well how to discriminate what looks like road signs)
The training took 95 minutes on my laptop (with its Nvidia 3060 GPU).
When applied on the all detected signs, the classification confidence :
- less than 90% in 4.5% of the time
- more than 99% in 87% of the time
Here the confusion matrix :
and an example of classification :
I published the model on Panoramax/classify_nl_road_signs · Hugging Face
The same process can be easily replicated on any country 
