AiSDF: Structure-aware Neural Signed Distance Fields in Indoor Scenes

Abstract

Indoor environments, where we commonly reside, often exhibit visual homogeneity or lack of texture, yet they inherently have structural characteristics that can serve as valuable priors for 3D scene reconstruction. Building on this observation, I introduce a structure-aware online reconstruction framework based on the signed distance field (SDF) tailored to indoor scenes, specifically under the Atlanta world (AW) assumption. I refer to this incremental SDF reconstruction method as AiSDF. In this online framework, I estimate the underlying Atlanta structure of the scene, followed by extracting planar surface elements with the surfel representation aligned with the estimated structure. These structure-aware surfels serve as an explicit planar representation of the scene. Furthermore, leveraging these planar surfel regions, I adaptively sample 3D points considering the complexity of the scene and enforce structural regularity during the SDF reconstruction process. This approach enhances reconstruction quality by preserving high-level structure while capturing scene details. I validate the effectiveness of AiSDF on the ScanNet and ReplicaCAD datasets, showcasing its ability to accurately reconstruct fine object details implicitly and represent structures explicitly in room-scale environments.