Hi everyone,
I recently got interested in the traffic sign classification problem, and came across 2 small datasets (the Belgian and rMASTIF (Croatian) traffic signs datasets) that seem to have been mostly ignored recently (I did not find any recent result on these datasets). I thought this could be a good opportunity to practice using the tools provided by fast.ai library and course, and see what accuracy could be obtained.
So this is basically what I did, and I must say that I obtained surprisingly good results. In particular, I got 99.4% test accuracy on the belgium traffic sign dataset, and did not find any better result publicly available (the best one I could find was 98.8%). My code is available on Github.
I would be very grateful for some feedback and comments on this work: did I miss something? Do you know some recent work on those datasets? What could be done to further improve these results? etc…
Thanks in advance !!