Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering
Organizes neural Gaussians on anchors and predicts attributes from the viewpoint to cut redundancy and improve generalization on complex scenes.
Authors / Team
Tao Lu · Researcher
Year
2024
Deep Dive
Vanilla 3DGS can sprawl redundant Gaussians to overfit training views. Scaffold-GS places anchors on a sparse scaffold, spawns local neural Gaussians, and predicts their attributes from viewing direction and distance inside the frustum, with growing and pruning on anchors. At inference it activates anchors in view and filters trivial Gaussians, reporting compact storage and better behavior on specular, textureless, or wide-baseline settings at comparable FPS.
What we learn
- 01
Explicit structural anchors pull unconstrained Gaussian growth back toward geometric priors.
- 02
View-conditioned prediction helps encode view-dependent effects without storing them per Gaussian.
Verbatim quote
"Neural rendering methods have significantly advanced photo-realistic 3D scene rendering in various academic and industrial applications."— source ↗
Tags
Links
Sources