Power Transform-Based Non-Uniform Quantization Framework for Scale Attribute Compression in 3D Gaussian Splatting
Hyeong-Gyun Yang, Dong-Ha Kim, Byung-Yoon Choi, Kwan‐Jung Oh, Jun Young Jeong, Gwangsoon Lee, Jae‐Gon Kim · IEEE Access · 2026
This paper proposes a power transform–based non-uniform quantization framework for efficient compression of the scale attribute in 3D Gaussian Splatting (3DGS) within the ongoing standardization of V-PCC Amendment 1 (Amd.1). Due to the exponential mapping from the logarithmic domain to the rendering domain, the scale attribute exhibits strong nonlinearity, making conventional uniform quantization inefficient for preserving perceptually important information. To address this issue, the proposed framework applies a power transform to reshape the distribution of normalized scale values, with the power parameter adaptively selected based on an R–D criterion. In addition, piecewise linear scaling (PLS) is incorporated as a refinement step to further improve bit allocation efficiency. Experimental results under the Gaussian Splat Coding (GSC) Common Test Conditions demonstrate that the proposed framework consistently outperforms conventional uniform quantization, achieving average RGB-PSNR BD-rate reductions of 7.3% and 6.4% for forward-facing and object-centric datasets, respectively. The proposed framework preserves rendering quality while maintaining compatibility with the existing V-PCC Amd.1 pipeline, providing an efficient and practical solution for scale attribute compression in 3DGS.