Abstract
Surface reconstruction from 3D point clouds is a critical task in computer vision, recently dominated by neural implicit representations. However, these methods often struggle with scalability due to the slow convergence of neural networks and high computational costs. We present GridPull, a "neural-free" approach that optimizes a discrete distance field directly on a 3D grid. By utilizing a specialized "pulling" loss and Total Variation regularization, GridPull achieves state-of-the-art accuracy while being significantly faster and more scalable than traditional neural methods, capable of processing point clouds with millions of points.
Introduction
Reconstructing smooth, high-fidelity surfaces from raw 3D point clouds is fundamental for applications in robotics, AR/VR, and digital twins. While neural implicit representations—which overfit a network to represent a distance or occupancy field—have set new benchmarks for quality, they face a significant bottleneck: scalability. The computational cost of calculating distances for thousands of query points against millions of surface points in every training iteration makes these methods impractical for large-scale scenes or dense scans. GridPull addresses this by replacing the neural network with a directly optimized grid-based representation.
Methods
GridPull introduces a discrete approach to learning implicit representations, where the learnable parameters are signed distance values at the vertices of a uniform 3D grid.
Discrete Distance Field and Pulling Loss
Instead of a network forward pass, the distance at any query point $q$ is computed via fast trilinear interpolation of the eight nearest grid vertices. To align this field with the input point cloud, GridPull employs a Pulling Loss. For a sampled query point $q$, the current field's gradient is used to "pull" $q$ toward the surface, resulting in a projected point $p$. The loss minimizes the distance between $p$ and its actual nearest neighbor in the point cloud.
Regularization and Efficiency
Because discrete grids lack the inherent smoothness of neural networks, GridPull incorporates a Total Variation (TV) Loss. This encourages neighboring grid vertices to maintain consistent values, ensuring a continuous and smooth field. To further boost efficiency, the method uses Bandwidth Optimization, updating only grid cells within a specific distance of the surface. Lightning-fast nearest neighbor searches are facilitated by a KD-tree.
Results and Discussion
GridPull was evaluated across diverse benchmarks, including object-level datasets (ShapeNet, FAMOUS, Thingi10k) and large-scale scene datasets (Matterport3D, KITTI).
Performance and Scalability
GridPull consistently outperforms recent neural-based methods like NeuralPull (NP) and Predictive Context Prior (PCP) in both speed and accuracy. On the FAMOUS dataset, GridPull achieved a mean reconstruction time of 201 seconds, significantly faster than NP (605s) and PCP (>3500s). Notably, GridPull successfully reconstructed a KITTI dataset frame containing 13.8 million points in approximately one hour—a task typically impossible for standard neural implicit methods.
Visual Fidelity
Visual comparisons on ShapeNet demonstrate GridPull's superior ability to capture fine geometric details and thin structures that are often smoothed over or missed by neural counterparts.
Conclusion
GridPull demonstrates that neural networks are not strictly mandatory for high-quality implicit surface reconstruction. By leveraging well-regularized discrete grids, the method achieves superior speed and detail. While performance is currently tied to grid resolution, future work involving adaptive structures like Octrees could further reduce memory overhead, pushing the boundaries of large-scale 3D reconstruction.
Summary updated