AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians

AI summary

Overview

Research area: Computer vision, specifically 3D surface reconstruction of indoor and urban environments from posed multi-view images.

Technical level: Advanced. The paper assumes familiarity with Gaussian Splatting, signed distance fields (SDFs), neural radiance fields, MLP-based implicit representations, TSDF fusion, and surface evaluation metrics (accuracy, completeness, F-score, chamfer distance). Authors are from the State Key Lab of CAD & CG, Zhejiang University.

Scope: The paper introduces AtlasGS, a framework that combines implicit-structured Gaussian Splatting with an "Atlanta world" structural prior (one gravity-aligned vertical direction plus multiple horizontal directions) to reconstruct smooth, detailed indoor and urban surfaces. It was published as arXiv:2510.25129v1 [cs.CV] on 29 Oct 2025.

What This Paper Is About

Man-made indoor and urban scenes contain large low-texture planar regions such as floors, ceilings, unadorned walls, and building facades, which are hard for image-based 3D reconstruction because there are few distinctive visual features to match. Existing geometric priors either lack global consistency (monocular depth/normal priors produce bumpy, view-inconsistent surfaces) or are too restrictive (the Manhattan-world assumption fails on urban buildings that are not mutually orthogonal, as the paper illustrates in Fig. 1 with a yellow-boxed building). At the same time, Gaussian Splatting-based methods are efficient but their discrete primitives create discontinuous, broken surfaces, while implicit SDF methods are smooth but computationally expensive and weak at capturing complex detail. The goal is a representation that is simultaneously smooth, detail-preserving, and efficient.

Key Contributions

1. An Atlanta-world guided implicit-structured Gaussian Splatting framework that targets smooth indoor and urban scene reconstruction while preserving high-frequency details and rendering efficiency.
2. A semantic Gaussian Splatting representation that predicts the semantic probability of low-texture regions — floor, ceiling, and wall — with the semantic attribute defined as z ∈ R⁴ (wall, floor, ceiling, others).
3. Structure plane regularization with learnable explicit plane indicators, comprising 3D global planar regularization (normal alignment plus planar constraints on Gaussian positions) and 2D local surface regularization (aligning normals derived from rendered depth with the plane indicators).
4. Extensive indoor and urban experiments demonstrating the best surface reconstruction quality both quantitatively and qualitatively against state-of-the-art implicit and explicit baselines.

Main Findings

• Indoor accuracy and completeness improve substantially. On Replica, AtlasGS reports Acc 2.25 cm, Comp 4.08 cm, Prec 93.18, Recall 82.22, and F-score 87.35, versus MonoSDF (4.14, 5.38, 75.50, 70.89, 73.08), ManhattanSDF (4.76, 5.59, 68.80, 66.40, 67.57), 2DGS (4.76, 6.34, 74.54, 65.37, 69.64), GSRec (4.90, 6.89, 73.32, 67.69, 70.37), Scaffold-GS (8.58, 11.27, 63.53, 54.91, 58.89), and DN-Splatter (16.97, 15.52, 32.67, 30.90, 31.75).
• Real-scene indoor results follow the same pattern. On ScanNet++ AtlasGS reports Acc 3.22, Comp 4.09, Prec 87.59, Recall 87.47, F-score 87.48; the strongest baseline there is ManhattanSDF at Acc 3.96, Comp 4.98, F-score 76.67, while 2DGS drops to F-score 21.71.
• On ScanNet, AtlasGS leads on geometry while staying between the implicit and explicit methods in speed. ScanNet results: Acc 3.62, Comp 3.93, Prec 80.31, Recall 75.85, F-score 77.98, 27 mins training, 70 FPS. For comparison, ManhattanSDF and MonoSDF take > 7 h with < 10 FPS (F-scores 67.25 and 71.21), Scaffold-GS takes 12 mins at 279 FPS (F-score 50.17), 2DGS 11 mins at 118 FPS (F-score 39.27), DN-Splatter 12 mins at 145 FPS (F-score 20.22), and GSRec 35 mins at 261 FPS (F-score 63.30).
• Urban reconstruction on MatrixCity improves by a large margin. AtlasGS reports Acc 0.022, Comp 0.034, and CD 0.028, compared with GaussianPro (0.102, 0.081, 0.091), GSRec (0.048, 0.175, 0.112), 2DGS (0.115, 0.098, 0.106), and Scaffold-GS (0.328, 0.303, 0.316). The paper notes units are not specified in the main text for this table.
• Novel view synthesis is competitive rather than uniformly best. On Replica AtlasGS reaches PSNR 39.58, SSIM 0.9756, LPIPS 0.0766, below 2DGS (41.59, 0.9823, 0.0464) but above GSRec (36.00, 0.9574, 0.1205) and ScaffoldGS (38.08, 0.9660, 0.0961). On ScanNet++ AtlasGS records PSNR 22.51, SSIM 0.8321, LPIPS 0.2517 — the best SSIM and LPIPS in that table, with PSNR below GSRec's 22.96 and DN-Splatter's 22.76.
• Every component in the ablation helps. On ScanNet, the full model achieves CD 3.77 and F-score 77.98, versus 2DGS with depth and normal priors at CD 12.68 and F-score 39.27. Removing the 2D local surface regularization gives CD 3.97 / F-score 75.52; removing both regularization terms gives 4.10 / 74.23; removing the normal prior gives 3.89 / 76.30; removing the depth prior gives 4.23 / 74.22.
• Baseline failure modes are named explicitly. ManhattanSDF and MonoSDF are smooth but lose fine detail (the paper cites a "lamp" example in Fig. 3) and show floor discontinuities; Scaffold-GS overfits under large lighting variation; 2DGS produces protrusions and holes in textureless regions such as sea and building facades; GSRec produces sparse reconstructions trimmed by Poisson surface reconstruction.

Methodology in Plain English

The method starts from 2DGS, which represents a scene as flat 2D "surfels" rather than 3D blobs and renders them with tile-based rasterization and Ray-Splat intersection, followed by alpha blending to produce color, depth, normals, and semantics.

Rather than optimizing each Gaussian independently, AtlasGS builds a sparse feature grid over SfM points. Each voxel stores a geometry feature, a semantic feature, offsets for K local Gaussians, and a shared scale factor, with voxel size 0.01 and feature dimension 32. Two small MLPs decode each voxel into K sets of Gaussian attributes (opacity, scale, rotation) and semantic attributes, while a view-direction-conditioned MLP decodes color. Because a voxel's decoder predicts neighboring Gaussians together, an update to one Gaussian influences its neighbors, which produces locally coherent surfaces while keeping the explicit Gaussians that preserve high-frequency detail.

Semantics are lifted from 2D to 3D. A pre-trained semantic segmentation model supplies 2D pseudo-labels, and the rendered semantic probability is supervised with cross-entropy; gradients are stopped through the rendering weights so that noisy semantic supervision does not distort geometry.

Structural priors then come from the Atlanta world assumption. The scene is approximated by a dominant horizontal plane (floor indoors, ground outdoors) plus multiple vertical planes (walls, buildings). Learnable explicit plane indicators are defined by a gravity direction plus distance offsets for floor and ceiling. These are initialized by extracting floor and ceiling points from either the semantic-lifted SfM points or the semantic Gaussians, fitting with RANSAC, and taking the fitted floor normal as the gravity direction. If the angle deviation or offset discrepancy exceeds a threshold, the indicator is reinitialized from the semantic Gaussians. For urban scenes the ceiling plane is omitted.

Two regularizers keep Gaussians aligned to those planes. The 3D global planar regularization enforces normal alignment (Gaussians in wall regions should have normals perpendicular to gravity; ceiling and floor Gaussians parallel) and a planar constraint pulling Gaussian positions onto their plane. Because Gaussian representations decouple position and orientation, orientation-only supervision cannot fix where a wall's Gaussians actually sit, so a 2D local surface regularization computes normals from the rendered depth map and aligns them with the plane indicators, weighting terms by semantic probabilities to soften segmentation errors.

Training combines photometric loss, monocular depth and normal priors, the two regularization terms, the semantic loss, plus the distortion and normal-consistency losses from 2DGS, with weights 0.25, 0.1, 0.1, 1.0, 100, and 0.05. The model runs for 40,000 steps, with voxel growth between steps 1,500 and 20,000 (gradient threshold 2e-4, pruned when opacity falls below 0.005), 3D global planar regularization starting at step 7,000 and 2D local surface regularization at step 20,000. Surfaces are extracted with TSDF fusion, and everything runs on a single NVIDIA 4090D GPU.

Why This Matters

Research impact. The paper argues that the standard structural prior for man-made scenes — the Manhattan world assumption — is too narrow for urban scenes where buildings are not mutually orthogonal, and positions the Atlanta world model as a necessary generalization. It also takes a different stance from prior hybrid works that simply overlay an implicit SDF field on Gaussian optimization (which the paper says compromises quality through mutual interaction), instead embedding implicit voxel grids inside the Gaussian framework as a smoothness regularizer. The reported ScanNet++ numbers show a large gap (F-score 87.48 vs 76.67 for the best baseline), which is the kind of result that invites follow-up replication and extension.

Real-world applications (as identified or implied by the paper):

• Digital twins of buildings and cities.
• Robotic navigation in indoor and urban settings.
• Augmented reality overlays that need accurate geometry.
• Urban-scale city reconstruction, evaluated on the MatrixCity dataset.

Industry relevance. The efficiency profile matters commercially: AtlasGS trains in 27 mins on ScanNet at 70 FPS, versus > 7 h and < 10 FPS for the implicit baselines, and the abstract claims reconstruction in under 30 minutes. That places it in a practical middle ground — slower than Scaffold-GS (12 mins, 279 FPS) and 2DGS (11 mins, 118 FPS) but far more accurate, and far faster than NeRF-style SDF methods.

Future Directions

1. Speed up training and rendering. The authors state as a limitation that their training and rendering are slower than previous Gaussian-based methods such as Scaffold-GS, 2DGS, and 3DGS; decoding all Gaussians during rendering is named as the cause of the slowdown relative to prior Gaussian approaches.
2. Reduce dependence on the Atlanta world assumption and the fixed semantic label set. The method relies on a pre-trained semantic segmentation model for a limited set of elements (wall, floor, ceiling, other); the authors propose extending to SAM and other geometry priors to supply planar information.
3. Broaden applicability beyond structured man-made scenes. Because the plane indicators assume one dominant horizontal direction and multiple vertical directions, scenes violating that assumption are outside the current method's design.
4. Broader evaluation. The paper reports results on four ScanNet scenes, seven Replica scenes, four ScanNet++ scenes, and four MatrixCity city blocks; expanding scene coverage and reporting units consistently for the urban metrics (not specified in the main text) would strengthen the evidence.

Target Audience

Researchers and graduate students working on 3D reconstruction, neural rendering, and Gaussian Splatting; practitioners building digital twins, AR, or robot navigation systems who need accurate geometry in textureless indoor and urban environments; and readers already familiar with 2DGS, NeRF/SDF methods, and standard surface evaluation metrics. Readers without background in differentiable rendering or implicit surface representations will find the method sections demanding.

Authors’ abstract

3D reconstruction of indoor and urban environments is a prominent research topic with various downstream applications. However, existing geometric priors for addressing low-texture regions in indoor and urban settings often lack global consistency. Moreover, Gaussian Splatting and implicit SDF fields often suffer from discontinuities or exhibit computational inefficiencies, resulting in a loss of detail. To address these issues, we propose an Atlanta-world guided implicit-structured Gaussian Splatting that achieves smooth indoor and urban scene reconstruction while preserving high-frequency details and rendering efficiency. By leveraging the Atlanta-world model, we ensure the accurate surface reconstruction for low-texture regions, while the proposed novel implicit-structured GS representations provide smoothness without sacrificing efficiency and high-frequency details. Specifically, we propose a semantic GS representation to predict the probability of all semantic regions and deploy a structure plane regularization with learnable plane indicators for global accurate surface reconstruction. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches in both indoor and urban scenes, delivering superior surface reconstruction quality.