TriaGS: Geometry-Consistent 3D Gaussian Splatting

• TriaGS is a geometry regularization method for 3D Gaussian Splatting that uses multi-view re-triangulation to enforce a globally consistent surface.
• It replaces photometrically optimized depth with robust consensus points from triangulation, achieving state-of-the-art mean Chamfer distance of 0.50 mm on DTU.
• The approach leverages self-supervised, differentiable SVD with a Geman-McClure loss to stabilize training and enhance multi-view geometric accuracy.

TriaGS, short for Triangulation-Guided Geometric Consistency for 3D Gaussian Splatting, is a geometry regularization method for 3DGS that addresses a central weakness of standard Gaussian Splatting: photometrically optimized reconstructions can render excellent images while remaining geometrically under-constrained. In TriaGS, the key corrective mechanism is a differentiable, self-supervised multi-view re-triangulation test: a rendered 3D surface point from a reference view is projected into a bundle of neighboring views, a robust consensus 3D point is recovered by multi-view triangulation, and the discrepancy between the rendered point and this consensus is penalized during training. The method does not rely on external MVS depth, learned correspondence models, or pre-matched image features, and it reports a mean Chamfer Distance of 0.50 mm on DTU, described as the best result among explicit Gaussian-based methods in its table (Tran et al., 6 Dec 2025).

1. Problem setting and geometric deficiency in vanilla 3DGS

Standard 3DGS models a scene as Gaussian primitives $G_i$ with mean $\boldsymbol{\mu}\in\mathbb{R}^3$, covariance $\boldsymbol{\Sigma}=RSS^TR^T$, opacity $\alpha$, and spherical harmonics coefficients for color. Rendering uses alpha compositing, and the rendered depth is written as

$$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$

Here, $\mathbf{u}$ is a pixel, $z_i$ is the depth of Gaussian $i$'s center, $\alpha_i'$ is that Gaussian's per-pixel opacity contribution, and $T_i$ is accumulated transmittance up to Gaussian $\boldsymbol{\mu}\in\mathbb{R}^3$0 (Tran et al., 6 Dec 2025).

The TriaGS formulation treats this standard depth as geometrically flawed because it blends depths of Gaussian centers rather than actual ray-surface intersections. If a Gaussian is slanted or disk-like, the ray’s true intersection with its density field does not generally occur at the center. The consequence is not merely local depth noise. The paper emphasizes a cascade of familiar failure modes: floaters, unstructured or “lumpy” geometry, inconsistency across views, and poor downstream surface extraction. Multiple semi-transparent Gaussians at different depths may reproduce the correct color without lying on a physically valid surface, so photometric optimization alone encourages image realism rather than geometric faithfulness.

A central claim of TriaGS is that global geometry should be enforced by multi-view consensus rather than only local pairwise checks. Prior Gaussian surface reconstruction methods are characterized as relying on local regularization such as planar constraints or pairwise depth or normal consistency. TriaGS argues that pairwise supervision is fundamentally limited because locally acceptable misalignments can accumulate across views and produce global geometric drift. Its alternative is to test whether a rendered 3D point is jointly consistent with many surrounding views at once.

2. Geometrically consistent depth, normal rendering, and 3D point construction

TriaGS begins from standard Structure-from-Motion (SfM) initialization. Camera poses are estimated, sparse SfM points initialize the Gaussian cloud, each camera $\boldsymbol{\mu}\in\mathbb{R}^3$1 has a shared intrinsic matrix $\boldsymbol{\mu}\in\mathbb{R}^3$2 and a rigid transform $\boldsymbol{\mu}\in\mathbb{R}^3$3, and the corresponding $\boldsymbol{\mu}\in\mathbb{R}^3$4 projection matrix is denoted $\boldsymbol{\mu}\in\mathbb{R}^3$5.

Because alpha-blended center depth is unreliable, TriaGS builds on RaDe-GS to obtain physically meaningful depth and normals. For a camera ray

$\boldsymbol{\mu}\in\mathbb{R}^3$6

RaDe-GS defines the ray-Gaussian intersection as the point of maximum Gaussian density along the ray, with

$\boldsymbol{\mu}\in\mathbb{R}^3$7

where $\boldsymbol{\mu}\in\mathbb{R}^3$8 is the camera origin, $\boldsymbol{\mu}\in\mathbb{R}^3$9 is the viewing direction, $\boldsymbol{\Sigma}=RSS^TR^T$0 is the Gaussian center, and $\boldsymbol{\Sigma}=RSS^TR^T$1 is the Gaussian covariance. The paper states that evaluating this directly per pixel is too expensive, so RaDe-GS uses the local affine approximation from 3DGS and a local ray-space simplification. From the resulting $\boldsymbol{\Sigma}=RSS^TR^T$2, depth is approximated by

$\boldsymbol{\Sigma}=RSS^TR^T$3

where $\boldsymbol{\Sigma}=RSS^TR^T$4 is the angle of the Gaussian center relative to the camera axis. Normals are computed in ray space and mapped back to camera space with

$\boldsymbol{\Sigma}=RSS^TR^T$5

Given a reference-view pixel $\boldsymbol{\Sigma}=RSS^TR^T$6, TriaGS renders its depth $\boldsymbol{\Sigma}=RSS^TR^T$7 and normal $\boldsymbol{\Sigma}=RSS^TR^T$8, then computes a current 3D surface point $\boldsymbol{\Sigma}=RSS^TR^T$9 by intersecting the viewing ray with the rendered local plane defined by depth and normal. The paper does not provide a separate closed-form ray-plane formula, but it states that $\alpha$0 is a differentiable function of Gaussian attributes because the rendered depth and normal are differentiable.

For each reference view, TriaGS selects a set of $\alpha$1 neighboring views using a deterministic neighbor view selection policy. For all headline results, it sets

$\alpha$2

The reference view plus these $\alpha$3 neighbors form the $\alpha$4 views that participate in the triangulation consistency check.

3. Differentiable multi-view triangulation and the robust consensus point

The correspondences used by TriaGS are not obtained by feature matching or photometric alignment. Instead, the currently rendered 3D point $\alpha$5 is projected into all selected views. For each view $\alpha$6,

$\alpha$7

where $\alpha$8 is the projection matrix, $\alpha$9 is the projected 2D observation, and $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$0 is the projective depth or homogeneous scale (Tran et al., 6 Dec 2025).

Given $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$1, TriaGS follows classical Direct Linear Transform triangulation. Each observation contributes two homogeneous constraints: $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$2 where $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$3 is the $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$4-th row of projection matrix $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$5, and $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$6 is the unknown homogeneous 3D point. Stacking all $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$7 views yields

$$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$8

with $$D(\mathbf{u}) = \sum_{i=1}^N T_i \alpha_i' z_i, \quad \text{where } T_i = \prod_{j=1}^{i-1}(1-\alpha_j').$$9. Because the system is generally inconsistent in practice, the method minimizes the algebraic error

$\mathbf{u}$0

It performs

$\mathbf{u}$1

and takes the right singular vector corresponding to the smallest singular value, namely the last column of $\mathbf{u}$2, as the consensus homogeneous point $\mathbf{u}$3. The paper states that this is implemented with differentiable SVD in PyTorch and batched over many matrices.

The re-triangulated point $\mathbf{u}$4 is the paper’s robust consensus point. The robustness is twofold: it aggregates evidence from $\mathbf{u}$5 views jointly rather than pairwise, and it uses a robust penalty on the residual $\mathbf{u}$6. The triangulation-guided loss is

$\mathbf{u}$7

This is the Geman-McClure form. The paper states that $\mathbf{u}$8 is annealed with an exponential decay schedule during training so that the loss becomes stricter in later stages. It also reports that a plain $\mathbf{u}$9 penalty causes optimization divergence and reconstruction failure, whereas Geman-McClure remains stable.

The paper repeatedly describes this mechanism as constrained multi-view triangulation. What is constrained is not triangulation by explicit inequality constraints; rather, the candidate point is required to satisfy the bundle of geometric constraints induced by all selected camera rays simultaneously. This distinguishes TriaGS from depth regularization, pairwise reprojection consistency, and methods that require external depth priors.

4. Training objective, optimization schedule, and self-supervised differentiability

TriaGS is not trained with the triangulation term alone. The intended total objective is described as

$z_i$0

where $z_i$1 is the primary photometric rendering loss, $z_i$2 is the triangulation-guided geometric consistency loss, $z_i$3 is a local geometric regularizer with normal consistency, and $z_i$4 is an auxiliary multi-view photometric consistency term with an occlusion-aware weighting. The paper notes that the typesetting is corrupted, but identifies this as the intended form (Tran et al., 6 Dec 2025).

The optimization schedule is explicit. The method is built on the official PyTorch implementation of 3DGS, uses RaDe-GS rendering for depth and normals, runs all experiments on a single NVIDIA RTX 4090, and trains for 30,000 iterations. The triangulation-guided geometric consistency term is enabled only after a 15,000-iteration warm-up, which the paper motivates by the instability of early geometry. Batched SVD is used to amortize the cost of many small triangulation systems. After training, the final mesh is extracted by TSDF fusion over rendered depth maps from all training views.

The process is described as both self-supervised and differentiable. It is self-supervised because the model itself predicts the rendered 3D point $z_i$5, projects it into neighboring cameras to create the observations $z_i$6, and then constructs an internal geometric consistency problem from those projections. It is differentiable because rendered depth and normal depend differentiably on Gaussian parameters, projection through $z_i$7 is differentiable, the matrix $z_i$8 is built differentiably, SVD is differentiable, and the residual-based loss is differentiable. A plausible implication is that gradients from $z_i$9 reach Gaussian means, covariances, opacities, and associated trainable parameters through the entire rendering-and-triangulation chain.

5. Benchmarks, ablations, and empirical behavior

The evaluation covers DTU, Tanks and Temples (TNT), NeRF-Synthetic, and Mip-NeRF 360 (Tran et al., 6 Dec 2025).

Dataset	Metric	TriaGS
DTU	Mean Chamfer Distance	0.50 mm
TNT	Mean F1-score	0.49
NeRF-Synthetic	Average Chamfer Distance	0.76
Mip-NeRF 360	Rendering quality	Competitive

On DTU, the reported means are: 3DGS 1.97, SuGaR 1.32, 2DGS 0.80, GOF 0.74, RaDe-GS 0.68, PGSR 0.53, Neuralangelo 0.61, and TriaGS 0.50. The paper highlights this as 5.7% better than previous explicit SOTA PGSR and 18% better than Neuralangelo. On Tanks and Temples, TriaGS attains a mean F1-score of 0.49, compared with 0.52 for PGSR, but the method is reported as faster, with 0.8h versus 1.2h. On NeRF-Synthetic, the average Chamfer Distance is 0.76, compared with 0.83 for PGSR, which the paper states as an 8.4% improvement. On Mip-NeRF 360, the claim is that TriaGS remains competitive on PSNR, SSIM, and LPIPS, especially on indoor scenes, showing that geometry regularization does not significantly degrade view synthesis performance.

The most informative ablation studies isolate the number of neighboring views $i$0 and the residual penalty. On Truck from Tanks and Temples, the reported values are: $i$1 gives PSNR 23.21, F1 0.57, 12.8 iter/s; $i$2 gives PSNR 24.11, F1 0.62, 11.6 iter/s; $i$3 gives PSNR 24.38, F1 0.66, 10.6 iter/s; and $i$4 gives PSNR 23.94, F1 0.71, 9.8 iter/s. The paper interprets this as evidence that more views improve geometric reconstruction, while pairwise consistency is inferior to multi-view consensus. Another ablation shows that Geman-McClure works stably whereas $i$5 causes optimization divergence and reconstruction failure.

The experimental setup is consistent with the method’s focus on geometry. DTU uses 15 object-centric scans as a surface reconstruction benchmark; TNT evaluates larger-scale scenes; NeRF-Synthetic tests synthetic detailed geometry; and Mip-NeRF 360 is used mainly for novel-view synthesis quality. This suggests that TriaGS is intended less as a compression or acceleration method than as a reconstruction method that tightens the coupling between rendered appearance and globally coherent surface geometry.

6. Position within the Gaussian splatting design space

TriaGS occupies the geometry-consistency corner of the recent Gaussian splatting design space. A nearby but distinct direction is efficient 3DGS. Trick-GS, for example, is not TriaGS itself; it keeps the standard 3DGS representation and proposes a carefully engineered combination of progressive training with resolution, noise and Gaussian scales, learning-based pruning and masking of primitives and SH bands, and an accelerated training framework, with the practical goal of improving storage, maximum Gaussian count, training speed, and rendering speed together (Armagan et al., 24 Jan 2025). TriaGS addresses a different bottleneck: the under-constrained geometry of photometrically optimized Gaussian splats.

Another adjacent direction is triplane-based GS generation. DirectTriGS represents Gaussian Splatting through a triplane field, decodes that field into geometry and Gaussian attributes, and targets 3D object generation rather than geometric consistency in multi-view reconstruction (Ju et al., 10 Mar 2025). Its central representational move is to replace a discrete Gaussian point cloud with a triplane-based continuous field. TriaGS does not change the underlying representation in that way; it remains within the standard explicit Gaussian reconstruction pipeline and strengthens it by enforcing multi-view geometric consensus.

This suggests a three-way distinction inside recent GS research. One branch targets resource-constrained deployment and compression, exemplified by Trick-GS. Another targets direct 3D generation through alternative latent representations, exemplified by DirectTriGS. TriaGS targets surface-faithful reconstruction by introducing differentiable multi-view triangulation as a geometry regularizer inside 3DGS training.

7. Limitations, failure modes, and future directions

The limitations stated for TriaGS are direct consequences of its reliance on multi-view geometry (Tran et al., 6 Dec 2025). The method is sensitive to camera pose inaccuracies, because triangulation depends on accurate multi-view geometry. It is also sensitive in sparsely viewed scenes, where neighboring views may be insufficient or weakly constrained, making the triangulation signal weaker. More generally, TriaGS is explicitly geometry-focused: it improves geometry, but its effectiveness depends on the availability of reliable multi-view evidence.

The paper also identifies a methodological constraint in the optimization itself. Early 3DGS geometry is noisy enough that a standard quadratic penalty on triangulation residuals is unstable, which is why the robust Geman-McClure form and the 15k warm-up are necessary. A common misconception would be to treat the triangulation loss as a drop-in replacement for pairwise consistency; the ablation results do not support that interpretation. The paper’s position is that pairwise agreement can remain locally acceptable while still drifting globally, whereas TriaGS solves one joint geometric agreement problem over many views at once.

The future directions proposed by the paper remain within that geometric framing. It suggests using triangulation residuals themselves for pose refinement and combining the method with hybrid implicit-explicit models to provide stronger priors where multi-view constraints are weak. Those directions are consistent with the method’s central thesis: the weakness of vanilla 3DGS is not its rendering efficiency, but the absence of a strong global mechanism requiring many views to agree on a common 3D surface point.