Hierarchical method improves real time 3d scene reconstruction quality and speed
Hierarchical Incremental Beta Splatting for Online Reconstruction
Computer Vision and Pattern RecognitionGraphics
Summary
Making 3D pictures of a scene while you capture it is tricky because early guesses about the shapes might be wrong and not fixed later. The authors provide a method that changes how parts of the 3D scene are created and adjusted, using a special shape that can be soft or solid depending on how confident the system is about the geometry. This approach helps keep solid parts where the shape is clear and soft parts where it's uncertain, and it recycles less useful parts to focus on detailed areas. Their method also speeds up the process and improves the quality of the 3D images as measured on a test set.
What this means in practice
- •For 3d scanning professionals: Improve real-time scene reconstruction quality and speed during live 3D capture workflows using hierarchical beta kernel splatting.$Commercial implications: Enables better and faster 3D scanning devices and software that produce higher-fidelity models in real time for industries like gaming and VR.
- •For computer graphics developers: Integrate confidence-aware primitives to refine online 3D models and optimize rendering via level-of-detail structures.
Authors
Linus Franke
Abstract
Online 3D Gaussian Splatting reconstructs a scene during capture: each keyframe spawns primitives once and refines them for only a few dozen iterations, leaving little chance to fix poor initialization or missing detail. Gaussians are also inherently soft, needing many overlaps for solid surfaces. We present Hierarchical Incremental Beta Splatting (HIBS), which brings beta kernels to incremental reconstruction. Their shape parameter interpolates between a soft, Gaussian-like profile and a flat-topped, near-opaque surfel, acting as a confidence dial for geometry. Our confidence-aware spawning uses geometric observability and multi-view depth agreement to initialize both opacity and shape, producing solid primitives where geometry is verifiable and soft, correctable primitives where it is not. As online densification, error-driven relocation recycles stale primitives into photometric-error hotspots. Bounded beta kernel support enables exact tile culling and faster rasterization, and a closed-form, shape-preserving merging formulation provides a level-of-detail (LOD) hierarchy for fast optimization and real-time rendering. On Mip-NeRF 360 in the on-the-fly protocol, HIBS improves PSNR by up to 0.8 dB while reducing optimization time by up to 40%. Project page: https://linus-franke.com/hibs/