HyperNeRF: Creating a dynamic 3D representation from a movie capture #WeeklyLabLink
Every Thursday at 11.55AM CET, we publish a link to a short, bite-sized online article, video or website that we think you should see.
This week's weekly lab link
Last week we explained how the makers of the Matrix movie managed to represent a 3D image from different angles using a huge amount of cameras, a big green screen and some clever computing power, and thus working around the limitation that classic Structure-from-Motion techniques only support static scenes. Their clever hack required a lot of hardware, so it is not available for consumers.
Recently a paper was published that might make it possible in the future to create a similar effect with just a regular smartphone: HyperNeRF.
A NeRF (Neural Radiance Field) is a Neural network that can predict what a location in space looks like from a certain angle for a specific static 3D scene. You can find a good explanation here. Unfortunately this only works on static scenes, so once you have a dynamic scene, NeRFs fail, as do most other Structure-from-Motion techniques
HyperNeRF found a way to work around this, as they added 2 extra dimension that also capture the movement in the scene. (Hyperspace is a term to define space that has more than 3 dimensions, hence the term HyperNeRF.)
While this technique certainly has its limits currently, it looks quite promising for the future: imagine having animated 3D scans available at the click of a button on your phone: you could watch a football match from the point-of-view you prefer, or look at a movie from another point-of-view...
Make sure to check out the paper's website as well, it makes the explanation accessible and has an interactive demo that shows you what the extra 2 dimensions in hyperspace represent.