WIP: Animation: Smoothing without Shrinkage #106365

Christoph Lendenfeld · 2023-03-31T10:20:04+02:00

2023-03-31 10:20:04 +02:00

This patch attempts to implement this paper for FCurve smoothing: https://web.archive.org/web/20190222092630id_/http://pdfs.semanticscholar.org/346b/fa0b3cdc11b3a6f1ef4e6a296b1edbd0c215.pdf

Thanks to Fabian Achammer for helping me understand what the paper says in the first place.

So the gist of it is:

FFT transform the data into its frequency components
apply this in frequency space
reverse the FFT

the main logic is in keyframes_general.c in case somebody is looking through the code

The FFT will be alright, since I am using the FFTW3 library.
Something must be wrong with the way I've implemented the logic in frequency space, most likely because I am not using the provided formula.

As far as I understand, the paper describes a frequency cutoff and a transition region.
Everything before the cutoff is unchanged, everything after the transition region is 0.
And everything in between lerps from 1-0, which is what I've done in code.

However the promise of no shrinkage isn't given as can be seen here

I think that also makes sense, if the high frequency tuples also have a high amplitude. But it might also be my limited understanding of the topic.

Now I am not sure if I can just apply a factor to the real and imaginary part of the FFT like I am doing.

This patch attempts to implement this paper for FCurve smoothing: https://web.archive.org/web/20190222092630id_/http://pdfs.semanticscholar.org/346b/fa0b3cdc11b3a6f1ef4e6a296b1edbd0c215.pdf Thanks to Fabian Achammer for helping me understand what the paper says in the first place. So the gist of it is: * FFT transform the data into its frequency components * apply this in frequency space ![image](/attachments/eb415968-9ea5-4287-80ea-748f2700a878) * reverse the FFT the main logic is in `keyframes_general.c` in case somebody is looking through the code The FFT will be alright, since I am using the FFTW3 library. Something must be wrong with the way I've implemented the logic in frequency space, most likely because I am not using the provided formula. As far as I understand, the paper describes a frequency cutoff and a transition region. Everything before the cutoff is unchanged, everything after the transition region is 0. And everything in between lerps from 1-0, which is what I've done in code. However the promise of no shrinkage isn't given as can be seen here ![image](/attachments/96ee30fb-98e0-4c78-b312-73db8852c797) I think that also makes sense, if the high frequency tuples also have a high amplitude. But it might also be my limited understanding of the topic. Now I am not sure if I can just apply a factor to the real and imaginary part of the FFT like I am doing.

image.png

68 KiB

image.png

21 KiB

Christoph Lendenfeld added 3 commits 2023-03-31 10:20:15 +02:00

boilerplate code 04982afd11

wip edfa580a1f

move code to keyframes_general.c e68307e5e7

Christoph Lendenfeld added 1 commit 2023-03-31 13:59:26 +02:00

curve length code 60bd26cb93

Christoph Lendenfeld added 2 commits 2023-04-01 21:13:25 +02:00

surprisingly working equal distance sampling fab90edf6d