PDRGS: Probabilistic Distribution Reshaping for Specular Geometry Reconstruction in 3D Gaussian Splatting

Abstract 3D Gaussian Splatting (3DGS) has shown remarkable capability for real‐time rendering and geometry reconstruction. However, reconstructing high‐quality geometry in specular scenes remains a persistent challenge, often leading to inward geometry collapse. We observe this inward collapse as a probabilistic distribution failure: cross‐view photometric variations cause the ray‐wise gaussian kernels distribution to become shifted and dispersed — a phenomenon we term peak drift and long‐tail distribution, which skews depth expectation away from the true geometry. To address this issue, we propose a probabilistic distribution reshaping framework for 3DGS that promotes both statistical stability and geometric consistency. First of all, we introduce an asymmetric spatial distillation strategy to prune incoherent long‐tail distribution, reshaping the dispersed probability mass back into a compact, surface‐aligned distribution. Building on this, we propose a distillation‐guided normal rectification strategy designed to anchor the drifting distributions to the true surface geometry. Finally, we incorporate a physics‐inspired neural shading module to disentangle complex appearance from underlying geometry, enabling normal‐aware gradient backpropagation. Extensive evaluations on standard benchmarks demonstrate that our approach yields lower errors, producing more compact Gaussian distributions and recovering fine geometric details that are challenging for existing approaches in highly specular scenes.