About ground truth generating

[amyburden]

I notice that you do score normalization in your paper. Can I ask about details about how you get the score normalization from l2 distance?

[javier-hernandezo]

Hello,

The test process is the next:

1) We extracted face embeddings for each image. 2) We computed the Euclidean Distance between images. These distances represent the dissimilarity between the pair of images. 3) The distances have been transformed to similarity scores s into the [0,1] range using the next equation: s = 1=(1 + e^d), being d each mated distance with zero mean and unitary standard deviation (we use a small sample of images to estimate the mean and the std of the distances previously).

Regards.

[amyburden]

@uam-biometrics Thanks you so much for such inspiring work. Use the face recognition model as supervision information is really a great idea. So I believe the equation should be s = 1/(1+e^d), right?

[amyburden]

@uam-biometrics Is it possible for you to provide image id in vgg2 face you choose as high-quality images?

[javier-hernandezo]

@uam-biometrics Thanks you so much for such inspiring work. Use the face recognition model as supervision information is really a great idea. So I believe the equation should be s = 1/(1+e^d), right?

Yes sorry, there was a mistake in my comment. s = 1/(1+e^d) is correct.

[javier-hernandezo]

@uam-biometrics Is it possible for you to provide image id in vgg2 face you choose as high-quality images?

Hello, I don´t have the list of images used as gallery images, but I will prepare it and upload it to the repository soon.

[javier-hernandezo]

The high quality images used for training have been uploaded to src/VGGFace2_Templates.txt

Regards