Alternating the filters about pytorch_wavelets HOT 4 CLOSED

Hi @CDitzel, thanks for the praise! Sorry it's taken me so long to get back to you.

This part of the code gets a little bit tricky, and I don't know if I can even remember this. I referenced Nick Kingsbury's original implementation a lot as the sample shifts are a little bit painful, especially with padding and remembering that the filters are themselves symmetrical.

These are the 2nd level filters, using the q-shift method: "A dual-tree complex wavelet transform with improved orthogonality and symmetry properties,” in 2000 ICIP.

I think the important part is figure 1

I've had a little look just now, but I can't work it out sorry. Perhaps contact Nick directly and he might be able to help out with the sample position logic a bit more?

from pytorch_wavelets.

@CDitzel I forgot to say, I also have another partly-tested implementation of the DTCWT based off of Selesnick's matlab implementation in pytorch_wavelets/dtcwt/lowlevel2.py:DTCWTForward2. You may find this one easier to follow than Kingsbury's one (the main one I use). I think I had difficulty getting the two to match up exactly because of the tricky things like sample shifts.

from pytorch_wavelets.

Hi Fergal, thank you for the detailed response. Yes I saw that, it follows a more traditional approach coding-wise. I will take a closer look. Did Professor Kingsbury introduce this rather obscure implementation that I was asking about?

from pytorch_wavelets.

Yes, it was written in an optimized way, but this makes it somewhat obscure

from pytorch_wavelets.