search

davidsandberg / facenet

Face recognition using Tensorflow
MIT License
13.83k stars 4.81k forks source link

Improve face alignment using a better align-dlib.py #19

Closed delip closed 8 years ago

delip commented 8 years ago

Per this new discussion on OpenFace, I think your models would benefit from a better dlib aligner. This is a drop in replacement for existing align-dlib.py. Checkout the new align-dlib.py.zip

davidsandberg commented 8 years ago

Very interesting thread! Definitely worth looking into. Thanks!

delip commented 8 years ago

FYI, this new alignment bumps accuracy from 0.9292 to 0.9425 on LFW

davidsandberg commented 8 years ago

That's a very good result! Did you just apply the new alignment to the LFW test set, or did you apply it to the training set and retrain as well?

delip commented 8 years ago

You need to retrain with the new alignments, and also apply to the test set. I would imagine not using the same alignment in test and train would make the results worse.

delip commented 8 years ago

Also, I did not do this experiment. This result is from Brandon @ OpenFace. I want to retrain the models to try out a bunch of ideas, but I'm not easily getting access to CASIA (I have Facescrub).

delip commented 8 years ago

Good news. I got access to CASIA dataset!