Improvement of Triple Attention about triplet-attention HOT 3 CLOSED

Hi, sorry for the late reply.

Can you talk about the inspiration for doing this, or explain it, I am confused about it

Sure. The main idea is to use regular convolution-based operations across three different "rotated" copies of the 3D input tensor. We hypothesized that this allows the network to capture cross-dimensional features in a way that regular 2D convolutions could not. For example, if the channel dimension represents time (such as in a video where frames are stacked and passed through a CNN), the network could learn features that correspond to an object moving up and down or side-to-side.

You can find more information about this idea of cross-dimensional interaction, the motivation for our method, and our analysis in our paper.

from triplet-attention.

Thank you very much for your full response！
What I mean is that this operation is actually a variant of spatial attention. So, have you considered doing the same on channel attention?

from triplet-attention.

What I mean is that this operation is actually a variant of spatial attention.

Yep, you're right. Triplet Attention applies a variant of CBAM's spatial attention across different permutations of the input.

So, have you considered doing the same on channel attention?

No, not really. But I suppose this would work as well, to a degree.

As we mention in the paper, one of the main motivations of Triplet Attention is efficiently learning how to extract "cross-dimentional" features. So by using spatial attention, each branch is able to learn features that span across two of the three dimensions.

If you used three channel attention branches, you wouldn't be able to learn those kinds of cross-dimensional feature extractors, since the canonical channel attention only operates across one dimension of a tensor.

from triplet-attention.