Points2Surf: Learning Implicit Surfaces from Point Cloud Patches (2007.10453v2)

Abstract: A key step in any scanning-based asset creation workflow is to convert unordered point clouds to a surface. Classical methods (e.g., Poisson reconstruction) start to degrade in the presence of noisy and partial scans. Hence, deep learning based methods have recently been proposed to produce complete surfaces, even from partial scans. However, such data-driven methods struggle to generalize to new shapes with large geometric and topological variations. We present Points2Surf, a novel patch-based learning framework that produces accurate surfaces directly from raw scans without normals. Learning a prior over a combination of detailed local patches and coarse global information improves generalization performance and reconstruction accuracy. Our extensive comparison on both synthetic and real data demonstrates a clear advantage of our method over state-of-the-art alternatives on previously unseen classes (on average, Points2Surf brings down reconstruction error by 30% over SPR and by 270%+ over deep learning based SotA methods) at the cost of longer computation times and a slight increase in small-scale topological noise in some cases. Our source code, pre-trained model, and dataset are available on: https://github.com/ErlerPhilipp/points2surf

Points2Surf: Learning Implicit Surfaces from Point Clouds

The paper "Points2Surf: Learning Implicit Surfaces from Point Clouds" introduces an advanced method for surface reconstruction that addresses the limitations of both classical and existing deep learning-based techniques. In scanning-based asset creation workflows, transforming unordered point clouds into a coherent surface is a critical task. Traditional approaches, like Poisson surface reconstruction, often falter in the presence of noise and incomplete data. Recent deep learning methods offer improvements but struggle with generalizing to new shapes that have varied geometrical and topological structures. Points2Surf presents a novel solution, integrating a patch-based learning framework that accurately reconstructs surfaces from point clouds without relying on normal vectors, leveraging both detailed local and coarse global information for improved generalization and accuracy.

Methodology

Points2Surf utilizes a dual strategy of global and local learning. The global model predicts the sign of the signed distance function (SDF), classifying whether a point is inside or outside a surface. Simultaneously, the local model uses a patch-based approach to calculate the absolute distance from the point cloud patches. This duality allows Points2Surf to maintain high reconstruction fidelity across a broad range of shapes and noise levels. The integration of local details, learned from the point cloud patches, with a global understanding of the shape structure, aids in retaining fine geometric characteristics while ensuring topological accuracy.

Results and Implications

The efficacy of Points2Surf is demonstrated through extensive testing on both synthetic and real-world data. Quantitatively, it achieves a substantial reduction in reconstruction error—30% when compared to Screened Poisson Reconstruction (SPR) and over 270% compared to state-of-the-art deep learning models like DeepSDF and AtlasNet. The qualitative analysis further supports this by showing its ability to preserve intricate geometric details and robustly handle topological variations.

Importantly, Points2Surf shows improved generalization capability across unseen datasets, which is a significant advancement over current deep learning methods that typically require retraining for each specific class of shapes. The ability to accurately reconstruct even complex and unseen shapes suggests broad applicability in various domains, such as gaming, augmented reality, and virtual reality, where the need for high-quality 3D models is prevalent.

Future Directions

This paper sets the stage for future exploration in multi-scale reconstructions where coarser levels ensure consistency for finer levels, potentially reducing surface artifacts and computational overhead. Moreover, enhancing the implicit function learning framework by integrating differentiable Marching Cubes could streamline the training process by allowing joint optimization of SDF estimation and surface extraction.

Conclusion

Points2Surf represents a significant methodological advancement in the domain of surface reconstruction from point clouds. By effectively combining local geometric detail with global shape analysis, it overcomes the limitations of both traditional and contemporary methods. Its robust performance across varying noise levels and ability to generalize to new shapes makes it a valuable tool in computational geometry and related applications. The paper provides a strong foundation for continued innovation in surface reconstruction techniques, promising both improved efficiency and higher fidelity in generating 3D models from raw data.