Radiance Fields
NeRF, MIP-NeRF, 3D Gaussian Splatting, and more
Intro
The concept of using 3D scanning as an alternative means for people to interact with their surroundings is intriguing. Museums could 3D scan their entire collections, capturing high-quality scans of artifacts and exhibits that can be shared with a global audience. This technology also enables the preservation of monuments at risk of being lost forever, allowing people to explore and learn from these cultural features remotely. Moreover, education can be enhanced by introducing more active learning methods based on working in 3D, which differing methods of constructing radiance fields make simpler.
Techniques such as NeRF, MIP-NeRF, 3D Gaussian Splatting, and others demonstrate a fast and simpler method of 3D scanning compared to traditional laser or projection scanning. These methods are not considered “production-ready.” These models are made up of very dense “point clouds” that cannot easily be translated into CAD. Supplementary software is getting better, but laser and projection scanning are still the best for high-accuracy work.
3D guassian spatting has been a new focus of research. You can use an entirely software-defined approach to create very light (the below model is 6 MB) and high-quality 3D models. With polygon-based models, you would have never been able to get this amount of accuracy in a 3D model that can run on a phone.
Showcase
3D Gaussian Splatting
Below you can see an example of a 3D guassian “splat” using the method Splatfacto, created by the engineers working on the Nerf Studio project inspired by the SIGGRAPH paper “3D Gaussian Splatting for Real-Time Rendering of Radiance Fields”.
left click rotate, right click pan
NeRF
Below are some pre-rendered photos based on the Nerfacto method also developed by nerfstudio. Though I have found that in some context the Splatfacto method produces a smaller file and is more efficient to run, NeRF provides pretty excellent quality.