Topic: DAIN (Depth-Aware Video Frame Interpolation)

I would love to see SVP integration with DAIN

Video with explanation what DAIN is:
https://www.youtube.com/watch?v=B1Dk_9k6l08

Source code:
https://github.com/baowenbo/DAIN

2 (edited by dlr5668 23-02-2020 08:08:51)

Re: DAIN (Depth-Aware Video Frame Interpolation)

10gb of vram for 720p smile

Re: DAIN (Depth-Aware Video Frame Interpolation)

I just came to post this video explaining the scholarly work after seeing it. Maybe the developers of SVP will find some of the information in it interesting.

4 (edited by dlr5668 23-02-2020 13:50:13)

Re: DAIN (Depth-Aware Video Frame Interpolation)

https://www.youtube.com/watch?v=r7IVm9FtFtQ
Doesnt look that good for me. Sepconv is better

5 (edited by Nintendo Maniac 64 24-02-2020 09:22:01)

Re: DAIN (Depth-Aware Video Frame Interpolation)

dlr5668 wrote:

10gb of vram for 720p smile

That's no issue for integrated graphics. tongue

...too bad DAIN currently only supports Nvidia GPUs, and of course that's the one of the "big 3" PC processor companies that doesn't make x86 CPUs which ruins my integrated graphics idea. sad


Alternatively, a CPU-only implementation may be an interesting experiment considering that we now have the likes of the Threadripper 3990X with it's crazy 64core/128thread configuration (and do remember that the $1000 i7-6950X from 2015 is now matched by today's ~$300 Ryzen 3700X, so that 3990X level of performance likely won't remain so out-of-reach for too long).


dlr5668 wrote:

https://www.youtube.com/watch?v=r7IVm9FtFtQ
Doesnt look that good for me. Sepconv is better

Really?  Personally I think DAIN looks considerably better in terms of having better and more natural movement as well as less-noticable and fewer artifacts, though with a big exception of choppier / faster-movement parts in which case Sepconv looks to have fewer artifacting from what seems to be less-aggressive interpolation on medium frame-to-frame changes.

To elaborate the last point, you can see in the following video at 0:42 that DAIN is poorly optimized for medium-level scene changes and (at least currently) doesn't do that sort of "repeat frame" thing that SVP does for such things unless it's a major scene-change:
https://youtu.be/Auum01OEs8k?t=42

Re: DAIN (Depth-Aware Video Frame Interpolation)

Heard about this a couple weeks ago. In my experience neural network trained algorithms tend to be slow and very marginal improvements in quality, but maybe someone here would know better.

https://grisk.itch.io/dain-app
https://github.com/baowenbo/DAIN#dain-d … erpolation
https://github.com/baowenbo/MEMC-Net