Reproduce “Compressing 3D Gaussian Splatting by Noise-Substituted Vector Quantization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haishan Wang, Mohammad Hassan Vali, Arno Solin
Researchers at Aalto University introduce NSVQ-GS, a method that applies Noise-Substituted Vector Quantization to 3D Gaussian Splatting, achieving an approximate 45x reduction in model size. This approach preserves high visual quality and real-time rendering speeds by enabling differentiable end-to-end optimization of Gaussian features and codebooks.
Key insights
- Noise Substitution in Vector Quantization (NSVQ) effectively resolves the gradient collapse problem inherent in VQ-based deep learning, allowing stable and differentiable optimization of discrete codebooks and continuous model features.
- Strategically quantizing specific Gaussian attributes (colors, SH, scale, rotation) while preserving others (coordinates, opacity) enables significant compression without introducing noticeable visual degradation.
- Joint end-to-end optimization of Gaussian parameters and codebooks with NSVQ can lead to compressed models that not only retain but sometimes even surpass the reconstruction quality (e.g., PSNR) of the original uncompressed 3DGS.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.