Introduction
Surface reconstruction—the transformation of discrete 3D point clouds into continuous solid meshes—is a fundamental challenge in computer vision. While existing methods like Poisson Surface Reconstruction (PSR) provide global consistency, they often sacrifice sharp details. Conversely, local methods like Floating Scale Surface Reconstruction (FSSR) handle varying scales well but lack the robustness of a global optimization framework. This paper introduces a variational approach designed to bridge this gap, capable of processing massive datasets of up to 1.5 billion points while maintaining high fidelity across multiple scales.
Real-world data presents significant hurdles: scale inconsistency between distant and close-up captures, sensor noise and outliers, and the sheer computational scalability required for billion-point datasets. The proposed method addresses these by integrating scale information directly into a global energy minimization framework and explicitly optimizing surface normals.
Methodology
The core of the approach is a global variational optimization performed over a balanced, scale-aware octree.
Scale-Aware Octree and Tetrahedral Mesh
The algorithm begins by constructing an octree where the depth adapts to the local scale of the input points. To simplify subsequent calculations, the tree is "balanced," ensuring that adjacent cells differ by no more than one level of depth.
Rather than optimizing directly on octree cells, the authors utilize a dual tetrahedral mesh. This structure allows for the application of Finite Element Methods (FEM) to solve the underlying equations with a minimal number of variables, significantly enhancing efficiency.
Energy Functional and Optimization
The surface is defined by minimizing an energy functional that balances four critical components:
- Data Fit (SDF): Alignment with input point positions.
- Data Fit (Normals): Alignment with input point orientations.
- Coupling: Consistency between geometry and normals.
- Smoothness: A regularizer that favors flat regions while permitting sharp discontinuities at edges.
By treating normals as primary variables rather than derived properties, the method preserves crisp edges and corners. Furthermore, the use of L1-like robust norms ensures the system remains resilient to noise and outliers.
Experimental Results
The framework was evaluated on several challenging benchmarks, including the Middlebury Temple, Citywall (256 million points), and the massive Breisach dataset (1.5 billion points).
Scalability and Accuracy
The reconstruction of the 1.5-billion-point Breisach dataset, covering 10,000 $m^2$, demonstrates the system's unprecedented scalability. Although the process required four days and 151 GB of RAM, it achieved a level of global consistency and detail previously unattainable at this scale. On the Middlebury Temple benchmark, the method consistently outperformed PSR and SSD in accuracy across multiple thresholds.
Qualitative Fidelity
Qualitative comparisons on the Citywall dataset highlight the method's ability to produce sharper corners and fewer artifacts compared to FSSR and PSR. The robust norms also proved effective in handling noisy data, such as reflections on car surfaces, resulting in cleaner reconstructions.
Discussion
The "secret sauce" of this approach lies in the explicit optimization of normals and the global nature of the solver. By considering the entire scene simultaneously, the algorithm can intelligently "fill in" missing data using surrounding context. To manage memory, the system employs compact histograms and k-means clustering within octree voxels, reducing the data term footprint to just 64 bytes per voxel.
While the computational cost is high—precluding real-time applications—the method sets a new standard for high-quality offline reconstruction. Future work will focus on optimizing solver speed and exploring more complex geometric priors to further enhance reconstruction quality.