Hacker News

points by cubefox 3 days ago | 0 comments

I suppose for version 3.0 they will abandon the whole green screen thing completely and the model will just recognize by itself (possibly with a user supplied hint mask) what the background is that needs to be removed and what the foreground is that should be kept, without any need for green screen or blue screen footage.

[–] sorenjan 3 days ago

Netflix has made a depth map model that can do pretty much that, although it's used for real time visualization not final production quality. https://eyeline-labs.github.io/FlashDepth/

[–] cubefox 3 days ago

Even at production quality, depth maps presumably don't work for keying transparent surfaces or transparencies from motion blur (both of which CorridorKey 2.0 is very good at) since a depth map doesn't create something like an alpha value. Though they may be helpful for creating somewhat coarse hint masks.