In some papers I've read that softmax loss is not preferred in FR since it does not give a good inter-class and intra-class margins, but could not understand 'why?'. So can someone explain, why softmax loss is not preferred in FR, in both mathematically and theoretically.
$\begingroup$
$\endgroup$
4
-
1$\begingroup$ Can you add some references (papers), so that people can follow your claim? $\endgroup$– PeterCommented Aug 16, 2019 at 11:20
-
1$\begingroup$ Recently I was reading a paper named "ArcFace" (arxiv.org/pdf/1801.07698.pdf), in which it was pointed out that the softmax loss is not prefered for FR and it gave references to many other papers for that claim, some of which are "SphereFace"( arxiv.org/pdf/1704.08063.pdf ) mentioned introduction chapter , next one is "Large-Margin Softmax Loss for Convolutional Neural Networks", this problem is highlighted in the beginning of the paper and also in the later sections of the paper. and there are many more references given by the "ArcFace" paper which claim the same. $\endgroup$– Manoj kumar ReddyCommented Aug 17, 2019 at 5:59
-
$\begingroup$ Please add the references to your question $\endgroup$– PeterCommented Aug 17, 2019 at 8:15
-
$\begingroup$ I have mentioned in the above already. $\endgroup$– Manoj kumar ReddyCommented Aug 17, 2019 at 10:29
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Disadvantage of softmax loss is written in Your referenced paper.
"ArcFace" (arxiv.org/pdf/1801.07698.pdf) and "Face recognition via centralized coordinate learning" https://arxiv.org/pdf/1801.05678.pdf
(1) the size of the linear transformation matrix W ∈ Rd×n increases linearly with the identities number n;
- there are millions of identities in the training data. Complexity will grow too much.
(2) the learned features are separable for the closed-set classification problem but not discriminative enough for the open-set face recognition problem.
- In an open-set problem, unknown classes may occur in the test stage. In a close-set problem, all test classes are known in the training stage. face recognition is open-set problem.
