Abstract
We introduce RGS-DR, a novel inverse rendering method for reconstructing andrendering glossy and reflective objects with support for flexible relightingand scene editing. Unlike existing methods (e.g., NeRF and 3D GaussianSplatting), which struggle with view-dependent effects, RGS-DR utilizes a 2DGaussian surfel representation to accurately estimate geometry and surfacenormals, an essential property for high-quality inverse rendering. Our approachexplicitly models geometric and material properties through learnableprimitives rasterized into a deferred shading pipeline, effectively reducingrendering artifacts and preserving sharp reflections. By employing amulti-level cube mipmap, RGS-DR accurately approximates environment lightingintegrals, facilitating high-quality reconstruction and relighting. A residualpass with spherical-mipmap-based directional encoding further refines theappearance modeling. Experiments demonstrate that RGS-DR achieves high-qualityreconstruction and rendering quality for shiny objects, often outperformingreconstruction-exclusive state-of-the-art methods incapable of relighting.