Abstract
While recent advances in Gaussian Splatting have enabled fast reconstructionof high-quality 3D scenes from images, extracting accurate surface meshesremains a challenge. Current approaches extract the surface through costlypost-processing steps, resulting in the loss of fine geometric details orrequiring significant time and leading to very dense meshes with millions ofvertices. More fundamentally, the a posteriori conversion from a volumetric toa surface representation limits the ability of the final mesh to preserve allgeometric structures captured during training. We present MILo, a novelGaussian Splatting framework that bridges the gap between volumetric andsurface representations by differentiably extracting a mesh from the 3DGaussians. We design a fully differentiable procedure that constructs themesh-including both vertex locations and connectivity-at every iterationdirectly from the parameters of the Gaussians, which are the only quantitiesoptimized during training. Our method introduces three key technicalcontributions: a bidirectional consistency framework ensuring bothrepresentations-Gaussians and the extracted mesh-capture the same underlyinggeometry during training; an adaptive mesh extraction process performed at eachtraining iteration, which uses Gaussians as differentiable pivots for Delaunaytriangulation; a novel method for computing signed distance values from the 3DGaussians that enables precise surface extraction while avoiding geometricerosion. Our approach can reconstruct complete scenes, including backgrounds,with state-of-the-art quality while requiring an order of magnitude fewer meshvertices than previous methods. Due to their light weight and empty interior,our meshes are well suited for downstream applications such as physicssimulations or animation.