Open emanuelconunaemme opened 4 years ago
@emanuelconunaemme with “native flutter” do you mean client side? How big do you forecast the model to be?
Sent with GitHawk
I think It will be quite hard to collect a large and robust dataset for this. Just to leave a pinpoint: https://arxiv.org/abs/1805.02283
Isn't better to integrate with SPID directly instead of handling all these id cards in the Rosseau system?
This Is the repo for previously mentioned Docface and Docface++ followup paper https://github.com/seasonSH/DocFace
@bhack yes, it's better to integrate spid. Although, this would cover 10% of the user. For the other 90% adopting some commercial solution (like tensorflow or aws rekognito) would filter 10% of the wrong id. I tried and it works pretty well. Of course, it's pretty inefficient in detecting the positives.
I still belive that with limited resources Is better to invest on Spid cause AI will not enable you for https://www.camera.it/leg18/126?tab=1&leg=18&idDocumento=1892&sede=&tipo=
So IMHO If we can do both It Is ok but if we need to choiche where to invest time probably it Is better to reserve resources for SPID.
We can integrate services like Rekognition, Tensorflow or native flutter AI libraries to filter non-IDs or bad pictures and ask the user to take another picture.