What are the differences between rot-inv-conv and G-Conv?

[a14en9]

Hi,

Thank you for sharing your code.

I just wander to know is there any differences between rot-inv-conv and Group "Equivariant Convolutional Networks" that has been proposed by Cohen?

Thanks!

[zhampel]

Hi @cvhabitat, I haven't read the paper, but clearly Cohen's approach is more general. In my implementation, I output the activation that is maximal given 90 or 45 deg rotated kernels. After briefly scanning the paper, it seems that the authors use all variants of the kernel (with some further generalization as well). If I have time to read it thoroughly, I can comment more specifically as well.

[a14en9]

Hi, Thank you for your reply. I compared with your code and Cohen's code. I think yours is much clearer and allows to rotate kernels in 45 degrees. So, I really want to understand the advantages of both methods. Thanks!

…

On Tue, 22 Oct 2019 at 17:41, Zigfried Hampel-Arias < ***@***.***> wrote: Hi @cvhabitat <https://github.com/cvhabitat>, I haven't read the paper, but clearly Cohen's approach is more general. In my implementation, I output the activation that is maximal given 90 or 45 deg rotated kernels. After briefly scanning the paper, it seems that the authors use all variants of the kernel (with some further generalization as well). If I have time to read it thoroughly, I can comment more specifically as well. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#13?email_source=notifications&email_token=AEIVGLUSNN5BLHKUI56UFLTQP4UNBA5CNFSM4JDP2QD2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEB6NBAQ#issuecomment-545050754>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEIVGLVU6KOQZ2MYF4UKOKTQP4UNBANCNFSM4JDP2QDQ> .

[a14en9]

Hi,

Have you tried to compare your idea with some baseline? For example, can you obtain better accuracy when you add such rotation invariance layer?

Best wishes,