MS${2}$IS: A Multi-Scale Semi-Implicit Surface Representation for 3D Reconstruction

Abstract

3D reconstruction is a key technology for recovering three-dimensional scenes from RGB-D data. Although mature solutions exist for 3D reconstruction based on RGB-D images, indoor scene reconstruction still faces challenges in balancing memory consumption, geometric representation accuracy, camera pose tracking accuracy, and reconstruction speed. To address these issues, we propose a multi-scale semi-implicit surface representation (MS<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula>IS) for 3D reconstruction. By leveraging multi-scale representation and adaptive voxel resolution control, our approach effectively balances memory consumption and geometric accuracy during the reconstruction process. Additionally, we maintain a gradient field within the multi-scale voxel structure, enabling the implicit voxel model to possess certain explicit representation capabilities, allowing for the direct extraction of surface vertices and normal vectors. This enhances both the accuracy and efficiency of camera pose tracking. Experimental results demonstrate that our method outperforms existing algorithms across multiple RGB-D datasets, achieving superior performance in both camera pose tracking and surface reconstruction.