Fast Position Bit Depth Estimation for Near-Lossless Gaussian Splatting Representation

Abstract

3D Gaussian splatting enables real-time, photorealistic novel view synthesis using millions of 3D Gaussian primitives, but its adoption is hindered by high storage demands. This letter presents a fast statistical method to estimate the optimal position bit-depth for near-lossless compression, without rendering or PSNR computation. By modelling duplicated point ratios in training data and applying outlier detection to test data, our method predicts the minimal acceptable bit-depth. Experiments on multiple datasets show that the method takes approximately 1.24 s on average. This performance is achieved while preserving near-lossless quality, making the approach practical for real-time and resource-constrained applications.