(2021-02-03 10:41)thag Wrote: Thankyou for this site, this has improved very much over the last few weeks, great programming.
Thanks, but its still not as fast as I would like it to be
(2021-02-03 15:39)metalmog Wrote: I've worked with face recognition once, and I can only imagine how large is your gallery. You talked about updating the database including new faces from last month. Does it work on an incremental way or do you retrain everything again? On my experience with face rec, updating the gallery frequently was a pain.
Yes it works incrementally because I do not use a neural network for the prediction. I ran exactly into this problem retraining would take too long.
I am sure there are ways around this, but I am also not an neural network expert.
For now the sites uses pretrained models for face detection and embedding/encoding. The prediction is then done by a nearest neighbor search.
(2021-02-03 15:39)metalmog Wrote: Another issue I had was regarding the image quality of query and gallery. Since the gallery was built with screencaps, I assume it's better if the query is also a screencap, instead of a profile picture from the model, correct?
Yes the site will work better if one also uses screencaps as search image, because the face encodings will be more similar.
(2021-02-03 15:39)metalmog Wrote: As a suggestion, I assume there is a lot of people inputting new images now, searching their favorite models. How about using them as well? You could have feedback from the users about the accuracy of the predictions. Something like, from the 10 closest models of the query submitted there is a checkbox confirming the correct models. In this way you could have new labeled faces (and bonus stats regarding which models are most looked, which websites, etc).
Since I do not use a neural network right now, there is less value in collecting user feedback. It would only be useful to detect if I screwed up something. Like a couple of days ago where only CB result where returned
Getting labeled training data is also not a problem.
Right now I download around 800k new images per day from the cam previews. At this point it is more about getting images from all models that exist, like new ones or models that are not online often. For all popular models I should have more then enough images.