For example, in PointNet, you see the 1D convolutions with the following channels 64 -> 128 -> 1024. Why not e.g. 64 -> 1024 -> 1024 or 1024 -> 1024 -> 1024?

  • $\begingroup$ @nbro It's not just PointNet, you see this design choice in pretty much all architectures. $\endgroup$
    – user3180
    Jul 17 '20 at 21:07
  • $\begingroup$ The number of kernels is typically associated with the number of different features that you want to detect in the input, although this is just an interpretation. So, in your case, if you increase the number of kernels (or channels), you are probably assuming that, as you process the input, more features should be detected. It would be nice if you could also provide other examples (just provide a link to the paper) where the number of channels or kernels increases as we go deeper, just for reference. $\endgroup$
    – nbro
    Jul 18 '20 at 16:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.