TriaGS: Differentiable Triangulation-Guided Geometric Consistency for 3D Gaussian Splatting

TL;DR

This work tackles geometry deficiencies in 3D Gaussian Splatting by enforcing global geometric consistency through differentiable multi-view triangulation. It introduces Triangulation-Guided Geometric Consistency (TGGC), which re-triangulates a consensus 3D point from multiple views and penalizes deviations from the rendered estimate, all in a self-supervised framework. The approach achieves state-of-the-art geometric reconstruction on benchmarks like DTU and Tanks and Temples, while maintaining competitive novel-view synthesis quality and offering a faster training workflow. These advances enable higher-fidelity surfaces suitable for downstream tasks requiring accurate geometry without extra supervision.

Abstract

3D Gaussian Splatting is crucial for real-time novel view synthesis due to its efficiency and ability to render photorealistic images. However, building a 3D Gaussian is guided solely by photometric loss, which can result in inconsistencies in reconstruction. This under-constrained process often results in "floater" artifacts and unstructured geometry, preventing the extraction of high-fidelity surfaces. To address this issue, our paper introduces a novel method that improves reconstruction by enforcing global geometry consistency through constrained multi-view triangulation. Our approach aims to achieve a consensus on 3D representation in the physical world by utilizing various estimated views. We optimize this process by penalizing the deviation of a rendered 3D point from a robust consensus point, which is re-triangulated from a bundle of neighboring views in a self-supervised fashion. We demonstrate the effectiveness of our method across multiple datasets, achieving state-of-the-art results. On the DTU dataset, our method attains a mean Chamfer Distance of 0.50 mm, outperforming comparable explicit methods. We will make our code open-source to facilitate community validation and ensure reproducibility.

Figures (9)

• Figure 1: TriaGS achieves superior surface reconstruction quality. Our method (b) successfully creates a smooth and detailed surface, addressing a trade-off in reconstructions from GS. TriaGS eliminates the "lumpy" surface artifacts seen in 3DGS (c) and SuGaR (e) while preserving finer details than over-smoothing methods like 2DGS (f), RaDe-GS (g), and PGSR (h). Our quality is achieved through global geometric consistency constraint, which is based on differentiable triangulation. Sub-figure (a) shows the rendered color and normal map.
• Figure 2: An overview of the TriaGS reconstruction pipeline. Our method begins with a standard SfM initialization (1). Our main contribution is a differentiable multi-view triangulation module (2) that enforces global geometric consistency. For a rendered point, we re-triangulate a robust consensus point from neighboring views. The residual between these points forms our novel Triangulation Loss, which guides the Gaussian Optimization (3). Finally, we extract the final high-fidelity mesh via TSDF fusion of rendered depth maps (4).
• Figure 3: Geometric inconsistency in standard 3DGS depth rendering. While matching the ground truth image (left) photometrically (middle), the standard depth map (right) is noisy and fails to capture the scene's geometry, making it an unreliable foundation for reconstruction.
• Figure 4: Limitation of Pairwise Consistency Local pairwise checks can accumulate errors, leading to global geometric drift. A check between views A and B may show only a "minor misalignment" and be considered consistent. However, this error compounds in subsequent checks (e.g., B to C), resulting in a significant "accumulated drift". This causes reconstructed surface (dashed red line) to deviate from the ground truth. Our method avoids this failure by enforcing a simultaneous multi-view consensus.
• Figure 5: Qualitative surface reconstruction on the DTU dataset. Our method (h) yields visibly superior geometric fidelity, which produces smoother, more complete surfaces with finer details. In contrast, methods relying on local consistency, like PGSR (g), still exhibit "bumpy" artifacts.