Abstract
Traditional surface reconstruction methods often struggle with open surfaces or produce noisy results due to the limitations of implicit distance functions. This paper introduces a grid-based intersection prediction approach that directly identifies where a surface crosses grid edges. By utilizing a Relative Sign Module and an Intersection Module with built-in symmetry constraints, the method significantly reduces artifacts and improves reconstruction quality for both watertight and open surfaces, achieving state-of-the-art results on benchmarks like ShapeNet and MGN.
Introduction
Converting raw point clouds into clean 3D meshes is a cornerstone of computer vision. Current deep learning approaches typically rely on Signed Distance Functions (SDF) or Unsigned Distance Functions (UDF). While SDFs are effective for watertight objects, they cannot represent open surfaces like clothing or sheets. Conversely, UDFs can represent open surfaces but suffer from gradient discontinuities at the surface boundary (the "gradient flip"), leading to jagged, low-quality meshes. This work proposes to bypass distance function regression entirely by directly predicting edge-surface intersections.
Methodology
The proposed pipeline shifts the focus from global distance field prediction to local intersection estimation. Instead of asking how far a point is from the surface, the model determines where exactly the surface crosses a specific grid edge.
Feature Extraction and Modules
The system uses a Point Transformer encoder to extract local geometric features from the input point cloud. These features feed into two primary components:
- Relative Sign Module: Determines if the endpoints of a grid edge lie on the same or opposite sides of the surface, providing the topological connectivity required for Marching Cubes.
- Intersection Module: For edges identified as crossing the surface, this module predicts the precise coordinate $\alpha \in [0, 1]$ along the edge where the intersection occurs.
Symmetry and Continuity
To ensure manifold continuity and prevent "cracks" in the mesh at grid boundaries, the authors enforce symmetry constraints. The modules are designed such that the predicted intersection remains invariant regardless of the order of the input point pair (e.g., $A \to B$ vs $B \to A$). This acts as a critical regularizer for producing smooth, continuous surfaces.
Results and Discussion
The method was evaluated across ShapeNet (watertight), MGN (open surfaces), and ScanNet (real-world scans) to demonstrate its versatility.
Performance Benchmarks
On the ShapeNet dataset, the approach achieved a 30% reduction in Chamfer Distance ($CD_1$) and a 1.8% improvement in Normal Consistency ($NC$) over previous state-of-the-art (SOTA) methods like GIFS. For open surfaces in the MGN dataset, the method produced noticeably cleaner edges and significantly fewer "floating" artifacts compared to traditional UDF-based approaches.
Efficiency and Robustness
Beyond accuracy, the method demonstrates superior efficiency. Inference for a ShapeNet object takes approximately 6 minutes, a substantial improvement over the 22 minutes required by the CAP method. Tests on ScanNet further confirmed the model's robustness when dealing with noisy and incomplete real-world data.
Conclusion
By reframing surface reconstruction as an intersection prediction problem, this method effectively addresses the gradient discontinuity issues inherent in UDFs. While challenges remain regarding GPU memory usage and boundary precision for open surfaces, the approach sets a new standard for high-fidelity 3D reconstruction from point clouds.
Summary updated