MagFace is a category of losses that learn a universal feature embedding whose magnitude can measure the quality of the given face, which is proposed in paper: MagFace: A Universal Representation for Face Recognition and Quality Assessment. In this tutorial, we will introduce it for beginners.
MagFace
This loss function is defined as:
Here \(a_i\) is the length of \(i\)-th face feature.
As to parameter \(u_m\), \(l_m\), \(u_a\) and \(l_a\), they are can be:
\(u_m = 0.8,
l_m = 0.45,
u_a = 100,
l_a = 10\)
As to \(\lambda_g\), it can be 20 or 35
Why we need to use \(\lambda_g g(a_i)\)?
As to \(g(a_i)\), when \(a_i = u_a\), it will get the minimum value, which means face magnitude can be \(u_a\) in maximum. In order to minimize \(g(a_i\)\), we need increase the value of \(a_i\).
Moreover, \(g(a_i)\) can make \(m(a_i)\) in \([l_m, u_m]\).