Introduction
High-fidelity 3D surface reconstruction from 2D images remains a fundamental challenge in computer vision. While recent advancements like 3D Gaussian Splatting (GS) have revolutionized rendering speeds, they often struggle to produce smooth, continuous surfaces, particularly in man-made environments. Conversely, implicit neural representations like Signed Distance Fields (SDFs) offer smoothness but at a significant computational cost. AtlasGS bridges this gap by introducing a hybrid framework that combines the efficiency of Gaussian Splatting with the structural consistency of implicit representations, guided by the "Atlanta-world" geometric prior.
The Challenge: Textureless Surfaces and Structural Trade-offs
Reconstructing indoor and urban scenes typically encounters two major hurdles:
- Textureless Surfaces: Standard reconstruction methods rely on distinct visual features to anchor geometry. In environments dominated by large, plain surfaces—such as white walls, ceilings, and floors—the lack of features often results in "bumpy," noisy, or incomplete geometry.
- The Efficiency-Smoothness Trade-off:
- Implicit Methods (SDFs/NeRFs): These produce continuous, high-quality surfaces but are computationally intensive and often lose fine details.
- Explicit Methods (3DGS): These are fast and detailed but consist of discrete primitives. Without strong constraints, these "blobs" can become disorganized, leading to "floaters" or disconnected surfaces in under-observed regions.
Methodology
AtlasGS addresses these issues through a two-pronged strategy designed to enforce local and global structural coherence.
1. Implicit-Structured Gaussian Splatting
To prevent the noise associated with independently optimizing Gaussian primitives, AtlasGS embeds them within a sparse feature grid. A lightweight Multi-Layer Perceptron (MLP) decodes the attributes of local Gaussians from this grid. This "implicit-structured" approach ensures that neighboring Gaussians are locally coherent and smooth, effectively inheriting the benefits of a continuous field while retaining the high-speed rasterization of the GS framework.
2. Atlanta-World Guided Regularization
While many models use the "Manhattan-world" assumption (where all surfaces are mutually orthogonal), this is often too restrictive for complex urban layouts. AtlasGS adopts the Atlanta-world model, which assumes a single vertical gravity axis but allows for multiple horizontal orientations. This is implemented via:
- Semantic Lifting: The model predicts the probability of a region belonging to specific classes like "floor," "ceiling," or "wall."
- Global Planar Regularization: Learnable plane indicators force Gaussians labeled as floors or ceilings to align with a global flat plane.
- 2D Local Surface Regularization: The model ensures that rendered normals in wall regions are consistent with the global gravity vector, preventing the protruding artifacts common in standard 2D Gaussian Splatting.
Experimental Evaluation
The researchers evaluated AtlasGS across several standard benchmarks for both indoor and outdoor environments:
- Indoor Scenes: ScanNet, ScanNet++, and Replica were used to test the model against real-world noise and high-quality synthetic meshes.
- Outdoor/Urban Scenes: MatrixCity, a large-scale synthetic city dataset, was used to evaluate the model's performance on complex building facades and diagonal street layouts.
Key Results
AtlasGS consistently outperforms state-of-the-art (SOTA) methods such as 2DGS and MonoSDF in both quantitative and qualitative metrics.
- Accuracy and Completeness: On the Replica dataset, AtlasGS achieved an F-score of 87.35, significantly higher than 2DGS (69.64) and MonoSDF (73.08).
- Efficiency: While the decoding step (MLP + feature grid) makes AtlasGS slightly slower than "pure" 3DGS, it remains an order of magnitude faster than implicit methods. A ScanNet reconstruction takes approximately 27 minutes with AtlasGS, compared to over 7 hours for MonoSDF.
- Surface Quality: Qualitative results demonstrate that AtlasGS eliminates the "broken" appearance of walls and floors, producing "water-tight" meshes that closely resemble ground truth, particularly on large-scale building facades.
Discussion and Limitations
The primary innovation of AtlasGS is the realization that semantic labels can serve as powerful geometric anchors. By knowing what an object is, the model can apply specific geometric priors (like planarity) to clean up low-texture regions. Furthermore, moving from Manhattan-world to Atlanta-world constraints provides the flexibility needed for modern urban architecture.
However, some limitations remain:
- Computational Overhead: The hybrid nature of the model means it is not yet "instant" compared to the original 3DGS.
- Semantic Dependency: The method relies on pre-trained semantic segmentation models (e.g., Mask2Former). Errors in segmentation can lead to incorrectly applied geometric priors.
- Planarity Assumptions: While more flexible than previous models, AtlasGS still assumes a degree of planarity and may struggle with highly organic, non-planar environments like caves or dense foliage.
Conclusion
AtlasGS represents a significant step forward in 3D surface reconstruction, successfully merging the speed of Gaussian Splatting with the structural integrity of implicit representations. By leveraging the Atlanta-world prior, it provides a robust solution for reconstructing complex, man-made environments with high fidelity and efficiency. Future work may involve integrating general foundation models like SAM to provide broader geometric priors without being limited to fixed semantic classes.
Summary updated