2. Training DeepID2

[MuaazBin]

Hello,
I have the computational resource.
I can train the siamese network if you are willing to share the database and script.
Sorry for writing as an issue, couldn't find any other way to reach you
Regards,
Muaaz Bin Sarfaraz
[email protected]

[happynear]

Don't try DeepID2 anymore. That algorithm is out-of-date.
You may be interested in SphereFace (https://github.com/wy1iu/sphereface), which is the current state-of-the-art face verification algorithm.

[MuaazBin]

Thanks will give it a go then. Is it comparable against deepface and facenet implementations https://github.com/cmusatyalab/openface.

[happynear]

As I remembered, this repo got LFW accuracy of about 93%, which is far below state-of-the-art algorithms. Don't try this code.

[happynear]

As a comparison, SphereFace got 99.3% on LFW and my NormFace (I would be glad if you read my paper and try my codes) got 99.2%.

[happynear]

When the accuracy is above 99%, the LFW dataset will no longer provide a fair comparison. So even though SphereFace's accuracy is similar with NormFace, the performance on other datasets, such as MegaFace, would be very different.

As I tested, SphereFace got ~73% and NormFace got only ~67% on MegaFace dataset. So please directly try SphereFace if you are new in this area.