AGS-Mesh: Adaptive Splatting & Meshing

• AGS-Mesh is an adaptive framework that augments Gaussian splatting with selectively filtered geometric priors to improve mesh estimation in indoor room reconstruction.
• It employs depth normal consistency and adaptive normal regularization to refine noisy smartphone depth and normal data, preserving structure and edge details.
• The scale-aware meshing strategy, using IsoOctree extraction, reduces artifacts and computational cost while enhancing novel-view synthesis quality.

to=arxiv_search.search 玩彩神争霸json {"query":"(Ren et al., 2024) AGS-Mesh Adaptive Gaussian Splatting and Meshing with Geometric Priors for Indoor Room Reconstruction Using Smartphones", "max_results": 5} to=arxiv_search.search 北京赛车开奖json {"query":"(Pan et al., 17 Apr 2025) Geometric adaptive smoothed aggregation multigrid for discontinuous Galerkin discretisations", "max_results": 5} to=arxiv_search.search 大发官网json {"query":"(Waczyńska et al., 2024) GaMeS Mesh-Based Adapting and Modification of Gaussian Splatting", "max_results": 5} to=arxiv_search.search 北京赛车pkjson {"query":"^^^^10^^^^", "max_results": 5} AGS-Mesh is an adaptive Gaussian Splatting and meshing framework for indoor room reconstruction using smartphone captures. It is designed for RGB images together with low-resolution mobile depth, and it augments Gaussian Splatting pipelines with geometric priors that are not used uniformly, but are filtered during optimization according to consistency. Its main components are joint surface depth and normal refinement, adaptive suppression of unreliable priors, and a scale-aware meshing procedure inspired by TSDF fusion and octree-based isosurface extraction. The method is reported for both 3D and 2D Gaussian Splatting-based pipelines, especially 3DGS/Splatfacto and 2DGS, with the stated goal of improving mesh estimation while preserving or improving novel-view synthesis on challenging indoor room datasets (Ren et al., 2024).

1. Reconstruction setting and motivation

AGS-Mesh targets indoor room reconstruction from smartphone data, including RGB imagery and low-resolution LiDAR depth from devices such as an iPhone Pro. The problem is difficult because smartphone depth is described as low-resolution and noisy, especially on edges, thin structures, and distant surfaces, while monocular depth and normal predictors can provide dense geometry cues but often have poor multi-view consistency and imprecise metric scale. Indoor scenes further combine large planar regions with small detailed structures, making fixed-scale reconstruction strategies suboptimal (Ren et al., 2024).

Within this setting, AGS-Mesh is positioned as a plug-in style enhancement for Gaussian Splatting pipelines rather than as a replacement for Gaussian Splatting itself. The method is motivated by the observation that purely photometric optimization in Gaussian Splatting often yields strong rendering but weak geometry in indoor scenes. Sensor depth supplies metric grounding on many smooth surfaces, and monocular priors can emphasize boundaries and object structure where phone depth is weak, but the priors are also inconsistent and region-dependent in reliability. AGS-Mesh addresses that inconsistency directly by filtering priors based on agreement during training rather than assuming they are globally trustworthy (Ren et al., 2024).

A central design choice is therefore not merely the inclusion of geometric priors, but the selective attenuation of those priors in ambiguous regions. The paper explicitly identifies object boundaries, thin structures, far-away surfaces, low-texture indoor surfaces, and regions where monocular predictors disagree with sensor geometry as cases where regularization should be reduced rather than enforced.

2. Depth Normal Consistency and depth refinement

The depth branch of AGS-Mesh is organized around a Depth Normal Consistency (DNC) strategy. The method first derives a robust normal estimate from the noisy sensor depth by backprojecting the depth map $D(x,y)$ into 3D, finding the $K$ nearest neighbors in world space for each depth point, computing the covariance of these neighbors, and taking the eigenvector corresponding to the maximum eigenvalue as a local normal estimate $\mathbf{N}_d$ (Ren et al., 2024).

That depth-derived normal is then compared to a pretrained monocular normal $\mathbf{N}_p$ using the angular discrepancy

$$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$

If the disagreement exceeds a threshold, the corresponding sensor depth is rejected: $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$ The rendered depth $\hat D$ is supervised by an $L_1$ loss, but the target changes over time: before the depth-filter activation step, the model uses the raw depth $D$; afterward it uses the filtered depth $D_f$ (Ren et al., 2024).

This mechanism makes the depth prior conditional on geometric self-consistency rather than on sensor provenance alone. In practical terms, the method is intended to preserve the metric usefulness of phone depth on smooth, reliable regions while preventing noisy or discontinuous depth values from dominating optimization near edges and thin structures. The paper’s ablations state that raw depth supervision already helps substantially, and that sensor depth supervision is much better than monocular depth supervision in this setting.

3. Adaptive Normal Regularization and optimization schedule

AGS-Mesh complements DNC with Adaptive Normal Regularization (ANR). The rendered normal is computed from the rendered depth field by

$K$0

This rendered normal is compared against the pretrained monocular normal prior, and the normal target is filtered according to angular agreement. When the angular discrepancy exceeds the threshold $K$1, the prior is zeroed out; otherwise the target remains the monocular normal $K$2. As with depth, the training schedule uses the raw prior early and the filtered prior later (Ren et al., 2024).

The supplementary schedule reported for AGS-Mesh is specific: $K$3k iterations for activating depth filtering, $K$4k iterations for activating normal filtering, total training of 30k iterations, thresholds $K$5, and loss weights $K$6 and $K$7. The RGB loss remains the original color objective of the underlying Gaussian Splatting model, described as $K$8 D-SSIM for a 3DGS-like pipeline (Ren et al., 2024).

The conceptual structure is twofold. First, AGS-Mesh checks prior-versus-prior consistency by comparing depth-induced normals with monocular normals. Second, it checks model-versus-prior consistency by comparing rendered normals with monocular normals. This staged filtering yields a warm-up regime in which the priors guide coarse structure, followed by a selective regime in which only mutually consistent estimates continue to regularize the scene representation. The qualitative outcome reported in the paper is improved planar smoothness, better edge preservation, reduced floaters, and more stable training.

4. Scale-aware meshing and IsoOctree extraction

AGS-Mesh does not stop at prior-regularized Gaussian optimization; it also introduces a scale-aware meshing strategy inspired by TSDF fusion and octree-based isosurface extraction. The motivation is that naive Marching Cubes on a uniform grid is expensive for large rooms, because fine voxel size is required to capture detail and the computational cost scales as $K$9. The method instead constructs a TSDF-like field from backprojected rendered depths and uses rendered normals to refine the contribution weights (Ren et al., 2024).

The isofunction given in the supplementary material is

$\mathbf{N}_d$0

where $\mathbf{N}_d$1 is the observed depth at the projection of point $\mathbf{N}_d$2 in view $\mathbf{N}_d$3, $\mathbf{N}_d$4 is the true depth of $\mathbf{N}_d$5 along camera $\mathbf{N}_d$6’s principal axis, and $\mathbf{N}_d$7 is the view weight. Truncation is depth-dependent, with a relative truncation parameter

$\mathbf{N}_d$8

Closer points receive a smaller truncation distance and farther points a larger one, with the truncation scaling linearly with depth (Ren et al., 2024).

Normal-aware weighting is implemented in two passes. The first selects a maximum-weight normal

$\mathbf{N}_d$9

with

$\mathbf{N}_p$0

The $\mathbf{N}_p$1 factor explicitly downweights distant and noisier measurements. The second pass refines the final $\mathbf{N}_p$2 using this normal information (Ren et al., 2024).

For surface extraction, AGS-Mesh uses IsoOctree rather than dense-grid Marching Cubes. The procedure starts from a uniform grid over the axis-aligned bounding box of the point cloud and subdivides voxels that contain sufficient point support. The paper specifies that a voxel is expanded if it contains at least $\mathbf{N}_p$3 points within radius $\mathbf{N}_p$4, with a maximum octree depth of 10. This yields a scale-aware hierarchy in which dense regions receive finer subdivision while sparse regions remain coarse, and the reported result is fewer artifacts and smaller meshes than standard TSDF plus Marching Cubes.

5. Experimental findings

The method is evaluated on real indoor mobile-capture datasets, principally MuSHRoom and ScanNet++. On MuSHRoom, mesh evaluation is reported for six scenes: coffee_room, honka, kokko, sauna, computer, and vr_room. The compared baselines include Volumetric Fusion, Nerfacto, Depth-Nerfacto, MonoSDF, 3DGS, SuGaR, GOF, Splatfacto, DN-Splatter, and the augmented variants 2DGS + Ours and Splatfacto + Ours (Ren et al., 2024).

For mesh quality on MuSHRoom, one of the strongest reported entries is 2DGS + Ours + SDF + IsoOctree, with Accuracy $\mathbf{N}_p$5, Completion $\mathbf{N}_p$6, Chamfer-$\mathbf{N}_p$7 $\mathbf{N}_p$8, Normal Consistency $\mathbf{N}_p$9, and F-score $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$0. The paper states that plain 3DGS, Splatfacto, and 2DGS without geometric priors produce weak meshes, while AGS-Mesh regularization improves Accuracy, Completion, Chamfer-$$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$1, Normal Consistency, and F-score. On ScanNet++, the method is likewise reported to improve mesh reconstruction quality, especially for 2DGS-based reconstruction.

For novel-view synthesis on MuSHRoom, the gains are also explicit. With 2DGS, the within-sequence test PSNR increases from $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$2 to $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$3, and the different-sequence test PSNR increases from $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$4 to $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$5. With Splatfacto, the within-sequence PSNR increases from $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$6 to $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$7. The paper emphasizes that the improvements are particularly notable for evaluation views from a different trajectory, which is presented as evidence of improved generalization and consistency (Ren et al., 2024).

The ablation study identifies the contribution of each component. On MuSHRoom, the reported 2DGS F1 values progress from baseline $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$8 to $$\theta_d=\arccos \left(\frac{\mathbf{N}_d \cdot \mathbf{N}_p}{\left\|\mathbf{N}_d \right\|\|\mathbf{N}_p\|}\right).$$9 with depth, $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$0 with both depth and normals, $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$1 with both plus DNC, $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$2 with both plus DNC and ANR, and $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$3 with both plus DNC, ANR, and IsoOctree. This establishes the intended contribution hierarchy: depth supervision supplies the largest initial gain, joint depth-normal supervision improves further, adaptive filtering adds additional gains, and the scale-aware meshing stage yields the best final mesh metrics.

Setting	Reported result
2DGS + Ours + SDF + IsoOctree	Accuracy 0.0249, Completion 0.0210, Chamfer-$$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$4 0.0229, Normal Consistency 0.8754, F-score 0.9146
2DGS within-sequence PSNR	22.52 $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$5 23.06
2DGS different-sequence PSNR	20.04 $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$6 20.97
Splatfacto within-sequence PSNR	24.47 $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$7 24.83
2DGS F1 ablation	0.6345 $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$8 0.8861 $$D_{f}= \begin{cases} 0 & \text{if}\ \theta_d > \tau_d \ D & \text{otherwise.} \end{cases}$$9 0.8880 $\hat D$0 0.9061 $\hat D$1 0.9092 $\hat D$2 0.9157

6. Relation to adjacent Gaussian-splatting methods and name ambiguity

AGS-Mesh belongs to a broader family of methods that couple point-based radiance-field representations to geometric constraints, but it occupies a distinct position within that landscape. It differs from "GaMeS" (Waczyńska et al., 2024), which parameterizes each Gaussian component by the vertices of a mesh face so that Gaussian position, rotation, and scale are directly induced by mesh deformation. It also differs from "DyGASR" (Zhao et al., 2024), which replaces ordinary Gaussian splats with generalized exponential splats, applies Generalized Surface Regularization and Dynamic Resolution Training, and then reconstructs a mesh by Poisson surface reconstruction.

This distinction matters because AGS-Mesh is not primarily a mesh-conditioned Gaussian representation in the GaMeS sense, nor a generalized exponential meshing pipeline in the DyGASR sense. Its central mechanism is adaptive filtering of depth and normal priors during optimization for smartphone indoor captures, followed by a scale-aware isosurface extraction strategy. A plausible implication is that AGS-Mesh is best understood as a prior-robust reconstruction and meshing framework layered onto existing Gaussian Splatting backbones rather than as a new primitive for Gaussian parameterization.

The term also carries a nomenclature ambiguity in current arXiv usage. A separate work on discontinuous Galerkin solvers uses "AGS-Mesh" to denote a geometric adaptive smoothed aggregation multigrid method for DG discretisations (Pan et al., 17 Apr 2025). That method is unrelated to smartphone room reconstruction and instead concerns mesh hierarchies, adaptive smoothed aggregation, SVD-based aggregatewise truncation, and near mesh-independent convergence for Poisson and convection-diffusion problems. In the 3D vision literature, however, AGS-Mesh refers to the adaptive Gaussian Splatting and meshing framework for indoor room reconstruction described above (Ren et al., 2024).