Research Milestone

Compact 3D Gaussian Representation for Radiance Field

Combines learned masking, grid-based view-dependent color, and codebook-quantized geometry to shrink Gaussian count and attribute storage while preserving quality.

Authors / Team

Joo Chan Lee · Researcher

Year

2024

Deep Dive

The paper tackles memory and disk growth of 3DGS as Gaussians densify. It learns a volumetric mask to drop low-impact Gaussians, replaces per-Gaussian spherical harmonics with a compact hash-grid color field, and codebooks geometric attributes where many Gaussians share similar scales and rotations. With quantization and entropy coding it reports large storage reductions and faster rendering versus vanilla 3DGS on benchmark scenes.

What we learn

01
Pruning redundant Gaussians and sharing attributes are the two main levers for compressing explicit radiance fields.
02
Moving high-frequency color from per-point storage to a continuous field cuts per-primitive parameters.

Verbatim quote

"Neural Radiance Fields (NeRFs) have demonstrated remarkable potential in capturing complex 3D scenes with high fidelity."— source ↗