Gaussian Haircuts?

Anyone who has ever tried to do volumetric capture of people will tell you that capturing hair can be a real pain. A head of hair is a complex, dynamic thing made up of thousands of strands (for most people anyway!), and volumetric capture methods can result in rendering the hair as a static, poorly defined blob - a bit like a Lego haircut!

What the researchers behind Gaussian Haircut have developed is a method for reconstructing hair as a collection of strands of gaussians. We've covered gaussians and gaussian splatting a few times before, but the basic idea is that a gaussian is a 2D ellipsoid that can represent a small part of 3D object, and a bit like pixels, if you have enough of them you can show the full object. Gaussians can also change colour based on what angle they're viewed, which means they can also capture things like reflections - very useful for glossy hair.

So a big part of what makes Gaussian Haircut so great is that it structures the hair gaussians so they align with strands, so the strand of hair is actually a strand of gaussians. Another cool thing is that this process results in outputs that can be taken into traditional 3D software and used for VFX and games. There's a demo on the project site where a hairstyle captured this way has been transplanted onto a metahuman character in Unreal Engine, and it looks great!