search

thunil / TecoGAN

This repo contains source code and materials for the TEmporally COherent GAN SIGGRAPH project.
Apache License 2.0
5.95k stars 1.13k forks source link

No improvement at all in the quality of the video #65

Open ziemowit-s opened 4 years ago

ziemowit-s commented 4 years ago

Hi, I tested TecoGAN on a video downloaded from YouTube: Jimmy Hendrix burning guitar, https://www.youtube.com/watch?v=3U5dvC5qr6Y

Input/output frames are: https://ibb.co/GWnsYMQ

My input frames are in 8bits, example file info: 0001.png: PNG image data, 320 x 240, 8-bit/color RGB, non-interlaced

Part of the original video after TecoGAN: https://youtu.be/YKVcHOHGSbQ

I see no difference at all. Do you know why?

caryknoop commented 4 years ago

When a source is too low quality, for instance when it has a lot of compression noise, this noise will only get amplified.

jahwanoh commented 4 years ago

does it mean that it is better to have a preprocess such as noise reduction or some kind of blur?

ziemowit-s commented 4 years ago

That's fine, but how to distinguish a good source?

Can you show an example of any old video enchanced with the TecoGAN (without the ground truth of the high resolution)?

For the other hand I wonder if the TecoGAN is a type of decompressor rather than quality enchancer?

Because as far I understand it have been trained on videos previously in high-res, downgraded to low-res. So it may just remember features of a particular compression method, rather than have the ability to enchance any video.

If you use different compression method, it probably fail, right?

caryknoop commented 4 years ago

Here is are two examples of Tecogan processed (among a lot of other things) SD interlaced videos: https://youtu.be/1ZwPe7YgiCY https://youtu.be/SKyrZiKVRvU

AloshkaD commented 2 years ago

@CaryKnoop the problem is there isn't a way to compare before-after to see how the algorithm works. In fact there is no mention that TecoGan was used for the restoration.

s-tweed commented 1 year ago

I'm also not seeing any improvement over a bicubic upsample on my end. I tried both the torch and tensorflow implementation hoping to see a good result with nothing to show for it. It does extremely well on the training sets, but nothing else I've tried shows much improvement at all. I've been able to get much better results just processing individual frames with real-ESRGAN than TecoGAN.

And I've definitely tried on SD footage with little to no compression/artifacting and that makes no difference.