The alternative of Centroid-Based Classification?

[ZhiyuanDang]

Thanks for your great work!

In order to do the final classification and regression, you propose a Centroid-Based method.

I think if it is possible, we first utilize the logarithmic map to map the final feature representations to the tangent space (i.e., Euclidean space) and then directly classify these by the FC and softmax layer.? ( Aim for classification task only.)

[leuchine]

Hi,

What you described is correct mathematically. I think it is just two different ways. What we used is more locality-aware in my opinion. But your method can still use the direction and norm information of features. I think both ways can be used up to the empirical results. Thanks!

Best Regards, Qi

[leuchine]

Another way is regarding our method as a kernel-motivated method, but your method is a softmax-based method. Does this make sense? Thanks!

[ZhiyuanDang]

Thanks for your reply.