Kerbl et al. (2023) — ARXIV-2308.04079
Key takeaways
- Three elements, all ablated separately: 3D Gaussians initialised from the sparse SfM points produced during camera calibration, interleaved optimization and density control including anisotropic covariance, and a visibility-aware rasterizer that accelerates training as well as rendering
- The key property is that one primitive has both: Gaussians are “a differentiable volumetric representation, but they can also be rasterized very efficiently by projecting them to 2D, and applying standard α-blending, using an equivalent image formation model as NeRF”. Prior work had to choose between optimisable and fast; this does not trade
- First real-time (≥30 fps) radiance-field rendering at 1080p for unbounded, complete scenes, at quality matching the expensive methods and training times competitive with the fast ones
- Static scenes. The optimisation referred to is density control during fitting; dynamic Gaussian work is later and separate
Standing in the anthology
Read — NOTE-045.
A source of SOTA-205, supplying the Gaussian primitives and the practice’s most portable corollary: initialise from structure you already computed. The SfM initialisation is ablated here (C2), which is the record’s only ablation of that claim.
It keeps NeRF’s (LIT-435) image formation and drops its representation. The paper says the Gaussians are rendered by α-blending “using an equivalent image formation model as NeRF”, so the volumetric radiance field and posed-image supervision carry over. The MLP queried along every ray is replaced by explicit anisotropic Gaussians that are projected and rasterized, which is where real-time 1080p rendering comes from.
The takeaway “dynamic scene optimization” was wrong. This paper is about static scenes — the abstract’s scope is “unbounded and complete scenes (rather than isolated objects)”, and the optimisation is density control of Gaussians during fitting. Dynamic 3D Gaussian work came afterwards. A reader taking that bullet at face value would cite this for something it does not do.