TY - JOUR
T1 - GPM-GS: 3D Gaussian Splatting Guided by Geometric Probability Models for Surface Reconstruction
AU - Wei, Xiaosong
AU - Zhou, Pengwei
AU - Ye, Shanding
AU - Zheng, Xiaoqi
AU - Li, Li
AU - Yao, Jian
PY - 2026/1
Y1 - 2026/1
N2 - Image-based 3D Gaussian Splatting (3DGS) lacks explicit geometric constraints, leading to inherent representational bottlenecks in surface reconstruction. This issue becomes especially pronounced in challenging scenes, such as those with weak texture and complex geometric structures. Although substantial progress has been achieved via monocular depth or multi-view geometric regularization, performance still degrades in complex scenes. The main reason is the insufficient exploitation of these valuable geometric features. To address this problem, this paper introduces GPM-GS, a novel approach that constructs a geometric probability model (GPM) using prior knowledge and self-supervision. This helps the Gaussian radiance field converge toward accurate surface representation. Specifically, we first establish a depth probability model (DepthPM). It guides the utilization of monocular cues through learnable uncertainty. These cues compensate for the limitations of photometric supervision, thus largely alleviating the inconsistency between 2D visual quality and 3D structural accuracy. Next, a calibration strategy is introduced based on a planar probability model (PlanarPM). It exploits local structural priors from monocular cues to correct under-reconstructed regions and improve geometric completeness. Finally, a cross-view probability model (CrossPM) is proposed to facilitate multi-view sampling and enforce cross-view consistency regularization. This maximizes the use of shared observations across views, effectively enhancing surface fidelity and fine-grained details. Extensive evaluations on various challenging scenes show that our method consistently achieves state-of-the-art (SOTA) reconstruction quality. © 1991-2012 IEEE.
AB - Image-based 3D Gaussian Splatting (3DGS) lacks explicit geometric constraints, leading to inherent representational bottlenecks in surface reconstruction. This issue becomes especially pronounced in challenging scenes, such as those with weak texture and complex geometric structures. Although substantial progress has been achieved via monocular depth or multi-view geometric regularization, performance still degrades in complex scenes. The main reason is the insufficient exploitation of these valuable geometric features. To address this problem, this paper introduces GPM-GS, a novel approach that constructs a geometric probability model (GPM) using prior knowledge and self-supervision. This helps the Gaussian radiance field converge toward accurate surface representation. Specifically, we first establish a depth probability model (DepthPM). It guides the utilization of monocular cues through learnable uncertainty. These cues compensate for the limitations of photometric supervision, thus largely alleviating the inconsistency between 2D visual quality and 3D structural accuracy. Next, a calibration strategy is introduced based on a planar probability model (PlanarPM). It exploits local structural priors from monocular cues to correct under-reconstructed regions and improve geometric completeness. Finally, a cross-view probability model (CrossPM) is proposed to facilitate multi-view sampling and enforce cross-view consistency regularization. This maximizes the use of shared observations across views, effectively enhancing surface fidelity and fine-grained details. Extensive evaluations on various challenging scenes show that our method consistently achieves state-of-the-art (SOTA) reconstruction quality. © 1991-2012 IEEE.
KW - 3D modeling
KW - 3D reconstruction
KW - Gaussian beams
KW - Gaussian distribution
KW - Geometry
KW - Image reconstruction
KW - Textures
KW - Three dimensional computer graphics
KW - 3d gaussian splatting
KW - Gaussians
KW - Geometric probability
KW - Geometric probability model
KW - Image-based
KW - Multi-views
KW - Probability modelling
KW - Regularisation
KW - Splatting
KW - Surfaces reconstruction
KW - Surface reconstruction
KW - 3D Gaussian Splatting
KW - Geometric Probability Models
KW - Surface Reconstruction
UR - https://openir.whu.edu.cn/AchievementDetail/Scopus_b1c441c6-de39-4917-b18b-606e14659e5b
U2 - 10.1109/tcsvt.2026.3731452
DO - 10.1109/tcsvt.2026.3731452
M3 - Article
SN - 1051-8215
JO - IEEE Transactions on Circuits and Systems for Video Technology
JF - IEEE Transactions on Circuits and Systems for Video Technology
ER -