QUVA-Lab / e2cnn

E(2)-Equivariant CNNs Library for Pytorch

Home Page:https://quva-lab.github.io/e2cnn/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

about to set special rotation equivariant

Newgaolujie opened this issue · comments

Thanks for your work!
According to my understanding,when i set r2_act = gspaces.Rot2dOnR2(N=8) ,it means that when the image is rotated by a 8 specific angle like ↑↗→↘↓↙←↖, the E2CNN outputs the same feature map ,in a similar fashion when N=4 ,the image is rotated by a 4 specific angle like ↑→↓←, the E2CNN outputs the same feature map.
I think it is possible to achieve the ←↖↑↗→ equivariant,but I'm not familiar with the source code. could you give me some hints?

Hi @Newgaolujie

Unfortunately, what you describe is not a group and, therefore, can not be achieved with our library (at least, not in a straightforward way).
As far as I know, this kind of problem has been approached by a few works in the literature, e.g. this paper. Maybe this can be a good starting point?

Best,
Gabriele

Thank you for your great help~