SolidGS: Consolidating Gaussian Surfel Splatting for Sparse-View Surface Reconstruction (2412.15400v1)

Abstract: Gaussian splatting has achieved impressive improvements for both novel-view synthesis and surface reconstruction from multi-view images. However, current methods still struggle to reconstruct high-quality surfaces from only sparse view input images using Gaussian splatting. In this paper, we propose a novel method called SolidGS to address this problem. We observed that the reconstructed geometry can be severely inconsistent across multi-views, due to the property of Gaussian function in geometry rendering. This motivates us to consolidate all Gaussians by adopting a more solid kernel function, which effectively improves the surface reconstruction quality. With the additional help of geometrical regularization and monocular normal estimation, our method achieves superior performance on the sparse view surface reconstruction than all the Gaussian splatting methods and neural field methods on the widely used DTU, Tanks-and-Temples, and LLFF datasets.

• The paper introduces a novel Gaussian kernel modification (SolidGS) that fuses splats into solid, opaque surface representations from sparse views.
• It integrates geometric constraints using virtual view losses and monocular normal estimations to guide accurate surface reconstruction.
• The approach significantly reduces training times while achieving state-of-the-art performance on benchmarks like DTU, Tanks-and-Temples, and LLFF.

Overview of "SolidGS: Consolidating Gaussian Surfel Splatting for Sparse-View Surface Reconstruction"

This paper introduces SolidGS, an innovative approach targeting sparse-view surface reconstruction in the field of 3D computer vision, utilizing Gaussian surfel splatting. The method addresses significant challenges in achieving high-quality surface reconstruction from limited input views by enhancing existing Gaussian splatting techniques.

Key Contributions

The SolidGS methodology presents several advancements over traditional Gaussian splatting techniques used in surface reconstruction:

1. Consolidated Gaussian Representation: The paper proposes a novel modification to the Gaussian kernel, termed as SolidGS, which integrates a learnable exponential factor across all Gaussian primitives. This adaptation enhances the opacity consistency of the Gaussians, ensuring they fuse into solid representations rather than diffuse, inconsistent ones. This consolidation is crucial in reducing depth inconsistencies, a prevalent issue in multi-view geometric rendering.
2. Geometric Regularization: SolidGS employs a set of geometric constraints to stabilize the optimization process. This includes geometry losses from virtual views and the incorporation of monocular normal estimations. These constraints are instrumental in guiding the optimization process towards accurate geometry, even with sparse input views.
3. Fast and Efficient Training: By rethinking the Gaussian splatting approach and employing effective initializations and constraints, SolidGS achieves higher fidelity reconstructions with significantly reduced training times. This efficiency makes it suitable for real-time applications like AR and VR, where quick processing is essential.

Numerical Results

SolidGS demonstrates substantial performance improvements across various datasets. In the DTU dataset, it achieves the lowest Chamfer Distance compared to other state-of-the-art methods, indicating superior surface accuracy. Interestingly, the training time is reduced to mere minutes, setting a new benchmark for efficiency in sparse-view scenarios. The approach also proves competitive in expansive and complex environments, such as those in the Tanks-and-Temples and LLFF datasets, where capturing geometric detail is notably challenging.

Implications and Future Directions

The research showcases significant potential for applications that rely on 3D surface reconstruction from limited data points, such as robotics, augmented reality, and remote sensing. SolidGS improves upon the scalability of Gaussian-based 3D modeling techniques, granting them broader applicability and effectiveness in diverse settings.

Moving forward, the paper’s approach could benefit from further exploration into enhancing real-time adaptability for dynamic scenes. Additionally, integrating SolidGS with other neural field techniques might yield even richer representations and renderings, thereby pushing the boundaries of what's achievable with sparse input data.

In summary, SolidGS presents a pioneering advancement in 3D surface reconstruction technology, offering enhanced accuracy and speed, solidifying Gaussian surfel splatting as a viable technique for real-world applications in computer vision.