nn4.v2 model released: LFW accuracy improved from 81.4% to 91.5%.

[Brandon Amos]

Hi OpenFace users,

I'm happy to release our nn4.v2 model, which I've been tracking progress on

in this GitHub issue: https://github.com/cmusatyalab/openface/issues/55

nn4.v2 uses the improved alignment technique that Hervé Bredin suggested

with improved training methods and slightly different model variations

by Bartosz Ludwiczuk.

On the LFW, the accuracy is improved from 81.4% to 91.5%, and

the ROC curve is:

The model is available from http://openface-models.storage.cmusatyalab.org/nn4.v2.t7 and

has md5sum 71911baa0ac61b437060536f0adb78f4.

Be careful using nn4.v2 with existing features extracted with nn4.v1 because

the features are not compatible.

The master branch now uses nn4.v2 by default, but should be fully

backwards compatible with nn4.v1.

The only parameter in the API distinguishing between these

two models is `landmarkIndices` during alignment.

This parameter defaults to openface.AlignDlib.INNER_EYES_AND_BOTTOM_LIP

for backwards compatibility with nn4.v1 and should be set to

openface.AlignDlib.OUTER_EYES_AND_NOSE for nn4.v2.

I've noted these differences at http://cmusatyalab.github.io/openface/models-accuracy/

More training images (over 500k) should further improve the network's accuracy.

Please contact me if you're able to collaborate with us and provide a dataset for training

an even more accurate model.

We're also looking into more sophisticated training techniques to improve the accuracy,

which are discussed in this mailing list thread:

https://groups.google.com/forum/#!topic/cmu-openface/dcPh883T1rk

-Brandon.