Impressive Video Data Compression

[anorlunda]

https://arxiv.org/pdf/2011.15126.pdf
https://nvlabs.github.io/face-vid2vid/
https://wandb.ai/ayush-thakur/face-...hesis-for-Video-Conferencing--Vmlldzo1MzU4ODc

One thing in this modern world seems to be ubiquitous; the demand for streaming more and more video. The
data compression in these papers, appears to be a significant step forward in reducing the bandwidth required.

I don't know whether to call it an algorithm or an AI. The difference is blurry.

On the dark side, it also appears to enable much simpler production of deep fakes.

 

[jedishrfu]

Interesting compression. Thanks for sharing.

This seems very similar to the recent animation of old photos using the old photo and a digitized actor doing the actions to be the guide for the video rendering.

They record an actor and key facial points for each frame and then map the old photo facial points to the actors facial points to render the scene.

I imagine too the artifacts can be ignored as side-effects of the video transmission in the viewers mind.

Standard video does something simpler with i-frames and p-frames where the i-frame is a full frame of the image (like jpg or bmp) and the p-frame is what changed.

https://en.wikipedia.org/wiki/Video_compression_picture_types