GVGS: Gaussian Visibility-Aware Multi-View Geometry for Accurate Surface Reconstruction

TL;DR

GVGS addresses the core challenge of accurate surface geometry recovery from Gaussian Splatting by introducing a Gaussian visibility-aware multi-view geometric constraint and a progressive quadtree-calibrated monocular depth constraint. The approach explicitly reasons about cross-view visibility at the Gaussian level and aligns monocular depth priors with Gaussian-rendered geometry through coarse-to-fine block-wise affine calibration, all within differentiable Gaussian splatting. Together, these components provide robust, stable geometric supervision, yielding state-of-the-art surface accuracy and mesh quality on DTU and Tanks and Temples, while also producing coherent visibility masks beneficial for broader 3D vision tasks. The work demonstrates that geometry-aware supervision can significantly enhance Gaussian-based surface reconstruction without sacrificing rendering efficiency.

Abstract

3D Gaussian Splatting enables efficient optimization and high-quality rendering, yet accurate surface reconstruction remains challenging. Prior methods improve surface reconstruction by refining Gaussian depth estimates, either via multi-view geometric consistency or through monocular depth priors. However, multi-view constraints become unreliable under large geometric discrepancies, while monocular priors suffer from scale ambiguity and local inconsistency, ultimately leading to inaccurate Gaussian depth supervision. To address these limitations, we introduce a Gaussian visibility-aware multi-view geometric consistency constraint that aggregates the visibility of shared Gaussian primitives across views, enabling more accurate and stable geometric supervision. In addition, we propose a progressive quadtree-calibrated Monocular depth constraint that performs block-wise affine calibration from coarse to fine spatial scales, mitigating the scale ambiguity of depth priors while preserving fine-grained surface details. Extensive experiments on DTU and TNT datasets demonstrate consistent improvements in geometric accuracy over prior Gaussian-based and implicit surface reconstruction methods. Codes are available at an anonymous repository: https://github.com/GVGScode/GVGS.

Figures (5)

• Figure 1: Overview of our GVGS. Given millions of 3D Gaussians, we render RGB images and depth maps from a reference view $v_r$ and a neighboring view $v_n$. We first obtain a depth-based weight by measuring the consistency between the rendered depth maps. To explicitly reason about cross-view visibility at the Gaussian level, we render the image of $v_n$ and compute per-Gaussian visibility weights $W_i$ according to Eq. (4). The visibility indicator of each Gaussian in view $v_n$, denoted as $\delta_i$, is then determined by thresholding $W_i > \tau$. Based on $\delta_i$, we render a Gaussian visibility-aware opacity $O_r$ using Eq. (5). Our final Gaussian visibility-aware weight is constructed by combining $O_r$ with the depth-based weight, and is used to enforce the Gaussian visibility-aware multi-view geometric consistency loss $L_{{gvmvgeom}}$. In parallel, for the ground-truth image, we predict a monocular depth map using Depth-Anything V2. This monocular depth prior is further aligned to the Gaussian-rendered depth through a quadtree-calibrated process guided by the Gaussian visibility-aware mask, and is finally used to formulate the quadtree-calibrated depth constraint $L_{{qdc}}$. Depth-based and Gaussian visibility-aware weights share the same color bar.
• Figure 2: Progressive quadtree-calibrated depth calibration. (a) Raw monocular depth prediction from Depth-Anything v2 yang2024depth and (e) the Gaussian-rendered depth, exhibiting clear depth bias and misalignment, especially in the roof regions on the right. (b--d) Depth maps after progressive quadtree-based calibration at increasing levels (Lv1--Lv3), where block-wise affine calibration is applied from coarse to fine. As the quadtree level increases, the calibrated monocular depth becomes progressively better aligned with the Gaussian-rendered depth.
• Figure 3: Qualitative ablation of the proposed components on Meetingroom. Red boxes highlight failure cases.
• Figure 4: Qualitative Comparison of Reconstructed Geometry with Related Works on DTU and Tanks and Temples. Red boxes highlight regions with noticeable geometric differences.
• Figure 5: Qualitative results of surface reconstruction and visibility estimation. (a) Surface reconstruction results of our GVGS on the DTU and Tanks and Temples datasets. (b) Qualitative comparison of visibility masks across different methods. The color visualization of DPFlow reflects 2D motion magnitude derived from optical flow, PGSR visualizes pixel-wise visibility weights inferred from photometric reprojection errors, while our method visualizes visibility weights obtained by aggregating the visibility of shared Gaussian primitives across views, producing cleaner and more coherent visibility.