Geo-3DGS: Multi-view Geometry Consistency for 3D Gaussian Splatting and Surface Reconstruction

Qingshan Xu, Wanjuan Su, Xinyao Liao, Xuanyu Yi, Jiequan Cui, Wenbing Tao, Yew-Soon Ong, Shuicheng YAN, Hanwang Zhang

Keywords: Gaussian Splatting, Multi-view geometry consistency, Surface reconstruction, Novel view synthesis

Abstract: Recently, the emergence of 3D Gaussian Splatting (3DGS) has made real-time and high-quality rendering possible. However, it is still challenging for 3DGS to reconstruct accurate geometry surfaces and achieve higher-quality rendering. To address these challenges, we propose to leverage multi-view geometry consistency for 3DGS and surface reconstruction. We reveal that there exists multi-view geometry inconsistency in 3DGS, preventing 3DGS from achieving higher-quality rendering and accurate surface reconstruction. To mitigate the geometry inconsistency, we first develop a multi-view photometric consistency regularization to constrain the rendered depth of 3DGS, which helps establish more stable and consistent 3D Gaussians to facilitate both rendering and surface reconstruction. To reconstruct geometry surfaces from 3DGS, we introduce a neural Signed Distance Function (SDF) field to represent continuous geometries of 3DGS. Then, we propose a geometry consistency-based SDF learning strategy, which leverages multi-view geometry consistency cues from 3DGS to efficiently optimize the SDF field for surface reconstruction. Extensive experiments on various datasets demonstrate that our method achieves both high-quality rendering and accurate surface reconstruction while keeping a good efficiency. Our code will be released upon publication.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9517