I don't think the result of the paper is worth it

[elk-april]

The affectnet dataset does not use any long-tail strategy, and it is basically impossible to achieve an accuracy of 65.20. However, this point is not mentioned in the paper. It can be considered that the author did not use this type of strategy. At the same time, considering that the author has no open source experimental settings of affectnet and difficult to reproduce, it can be determined that the experimental results are not trustworthy.

[zyh-uaiaaaa]

It seems the attentive center loss branch has little impact on the accuracy.