Immersive Light Field Video with a Layered Mesh Representation

They could probably cut it down significantly by only streaming a several-inch-sized sphere of free movement, predicting what segment to stream next. There is a limitation to how fast human beings can move/rotate their heads.

Bandwidth is getting cheaper over time. Couple of years ago I got a 56K modem, and now 1gbps connections are getting mainstream.

On a flatscreen I would say it's just a gimmick. With a VR headset though, since you are constantly moving your head a little I would say it's a massive improvement over 180° 3D videos, and really helps with the feeling of "being there".

20-60x the bandwidth requirements is one of the things that can reasonably be expected to improve with effort. Look at 2D video and how much better h264 is than, say, GIF. And the higher the dimensionality, the more data but also the more low-hanging fruit for compression. We had 100:1 lossy lightfield compression 20 years ago; we can likely put another zero on that, maybe more with machine learning.

That's what the linked paper is about - effective representations for lightfield data. The field is just kicking off.

Have you tried Welcome to Lightfields?

If they get that working with video, it's absolutely worth it.

Postproduction on this would be a nightmare though, so I wouldnt bet this being used on movies soon.

They did mention that they are compression the image data with H.265, so depending on the quality level, low-contrast areas might get smeared out. Also, there are some depth resolving errors here and there, which might be solved by using better algorithms. Or this might be an artifact from compressing the layer geometry (they used the Draco library, but did not specify in the paper what parameters they used). The paper mentions some of the limitations they encountered.