Why not use the softmax to obtain the MLR score?

[kinredon]

Thanks for this interesting work.

This paper uses a normalized score with the sigmoid function to calculate the MLR score.

Why not use the softmax to obtain the MLR score?

[frh23333]

Hello, this is because one image may have multiple categories. If we use the softmax function, the probability scores will become a one-hot distribution. For example, if there are 2 categories in the image, we hope the output should be like (0.01, 0.01, 0.99, 0.99). But softmax will make it to be like [0, 0, 0.5, 0.5].

[kinredon]

Thanks for your reply. I got this.

By the way, one of the baselines of the paper is BoxSup. But in the paper, it is hard to find its reference. Can you share the paper information with me? Thanks!

[frh23333]

BoxSup is a strong baseline model proposed in the paper of Detic, you can find the detailed introduction in their paper: https://arxiv.org/abs/2201.02605