TopoSurfel: Closing the Loop between Gaussian Surfels and Meshes for Surface Reconstruction
TopoSurfel introduces a differentiable mesh extraction and mesh-guided surfel evolution, achieving high-precision surface reconstruction with Chamfer Distance 0.70M and F1-Score 0.64, outperforming baselines.
Key Findings
Methodology
TopoSurfel integrates Gaussian surfels with a continuous proxy mesh via a non-trainable differentiable Marching Cubes (DiffMC). The pipeline begins with initial surfel optimization guided by photometric and geometric constraints, followed by normal alignment and density regulation based on mesh proximity. A differentiable Poisson surface reconstruction (DPSR) converts a weighted oriented point cloud into a scalar field, from which DiffMC extracts a continuous mesh serving as a global geometric prior. This mesh guides surfel evolution through normal alignment and geometry-aware density control, forming a closed-loop optimization. For large scenes, a hybrid re-initialization strategy based on scene-aware spatial cues ensures robustness. The entire process is end-to-end differentiable, enabling joint optimization of surfels and mesh, leading to accurate, artifact-free surface reconstructions.
Key Results
- On multiple benchmarks, the method achieves a Chamfer Distance of 0.70M and F1-Score of 0.64, surpassing MILo (CD=1.22, F1=0.59). It reconstructs detailed surfaces efficiently, with a runtime of 35 minutes for large scenes, significantly faster than previous methods.
- The mesh-guided surfel evolution effectively suppresses floaters and fills holes, especially in textureless or occluded regions, leading to more coherent and complete surface models.
- The hybrid re-initialization enhances stability in complex environments, enabling consistent high-quality reconstructions across diverse scenes, demonstrating the method's robustness and scalability.
Significance
This work bridges the gap between unstructured Gaussian surfels and explicit meshes, enabling high-fidelity surface reconstruction with global topological consistency. It addresses longstanding issues of artifacts and floaters in point-based methods, providing a scalable, efficient framework suitable for large-scale scene modeling. The integration of differentiable mesh extraction with surfel optimization opens new avenues for real-time, high-precision 3D reconstruction, impacting applications from virtual reality to industrial inspection. Its ability to produce detailed, artifact-free meshes while maintaining rendering quality marks a significant advancement in 3D vision and graphics.
Technical Contribution
The core innovation lies in the differentiable proxy mesh extracted via DiffMC, which provides a global geometric prior during surfel optimization. The mesh-guided surfel evolution strategy, including normal alignment and geometry-aware density control, ensures structural coherence. The hybrid re-initialization scheme stabilizes large-scale reconstructions. Unlike prior methods relying on local or post-processing steps, this approach enables end-to-end joint optimization, improving geometric fidelity and efficiency without additional learnable parameters, thus offering a new paradigm for point-to-mesh fusion.
Novelty
This is the first work to implement a fully differentiable, closed-loop system that tightly couples Gaussian surfels with a continuous proxy mesh via DiffMC. It introduces a mesh-guided surfel evolution strategy based on explicit geometric rules, significantly improving structural accuracy and artifact suppression over existing local or post-processing methods. The hybrid re-initialization further enhances large-scale scene robustness, setting a new standard in surfel-mesh integration.
Limitations
- Despite improvements, the method still faces challenges in scenes with extreme textureless regions or severe occlusions, where initializations may be less reliable.
- Computational complexity remains high, especially during the mesh extraction and re-initialization stages, limiting real-time applications.
- Parameter sensitivity, such as thresholds for normal alignment and density control, can affect stability and accuracy, requiring careful tuning.
Future Work
Future directions include integrating learned priors for better initialization, optimizing the pipeline for real-time performance, and extending the framework to dynamic scenes. Combining with neural implicit representations could further enhance detail recovery and topological robustness, broadening the applicability to robotics, AR/VR, and large-scale scene understanding.
AI Executive Summary
Surface reconstruction remains a fundamental challenge in 3D vision, especially when balancing detail, efficiency, and topological correctness. Traditional methods like Marching Cubes and TSDF fusion excel in controlled settings but struggle with complex, large-scale scenes, often producing artifacts or incomplete surfaces. Recent advances in neural implicit representations, such as NeRF and NeuS, offer high-quality rendering but lack explicit geometric control, limiting their use in precise modeling.
Gaussian Splatting (3DGS) introduced an explicit, point-based volumetric representation that enables real-time rendering and scalable scene modeling. However, its unstructured nature hampers direct extraction of continuous, high-fidelity surfaces. Existing reconstruction approaches either rely on post-processing or local geometric constraints, which are insufficient for ensuring global topological consistency and artifact-free surfaces.
To address these issues, this paper proposes TopoSurfel, a novel framework that tightly couples Gaussian surfels with a continuous, differentiable proxy mesh. The core innovation is the use of DiffMC, a differentiable Marching Cubes algorithm, which converts a weighted oriented point cloud—derived from surfels—into a continuous scalar field. This scalar field is then used to extract a proxy mesh that encodes global geometric priors.
The framework proceeds in two stages: a warm-up phase stabilizes the surfels using photometric and local geometric constraints, followed by a hybrid re-initialization that leverages coarse TSDF meshes for large scenes. During the main optimization, the proxy mesh guides surfel evolution through normal alignment and geometry-aware density control, suppressing floaters and filling holes. This closed-loop process ensures consistent, artifact-free surface reconstructions with high geometric fidelity.
Experimental results demonstrate that TopoSurfel achieves state-of-the-art accuracy, with Chamfer Distance as low as 0.70M and F1-Score of 0.64, while significantly reducing reconstruction time to 35 minutes for large scenes. The method effectively handles textureless and occluded regions, producing detailed, topologically consistent meshes suitable for applications in VR, digital twins, and industrial inspection. Despite some computational costs and parameter sensitivities, the approach marks a substantial step forward in scalable, high-quality 3D surface reconstruction, opening avenues for future integration with learned priors and real-time systems.
Deep Dive
Abstract
3D Gaussian Splatting has achieved remarkable success in novel view synthesis. However, extracting high-fidelity surfaces directly from 3DGS remains challenging due to its discrete and unstructured nature. Existing 3DGS-based reconstruction methods typically rely on multi-view geometric consistency or local constraints. Without an explicit structured geometric prior during optimization, these methods often struggle to resolve structural ambiguities, leading to artifacts and floaters, particularly in textureless or occluded regions. To address this limitation, we propose TopoSurfel, a novel framework that closes the loop between Gaussian surfels and continuous meshes. Unlike recent methods that incorporate mesh extraction into the differentiable pipeline by introducing auxiliary neural networks or extra per-Gaussian parameters, we dynamically extract a continuous proxy mesh via a non-trainable differentiable iso-surfacing process. Leveraging this differentiable connection, we introduce a mesh-guided surfel evolution strategy, including normal alignment and geometry-aware density control, to effectively suppress floaters and fill surface holes. Furthermore, to address the initialization challenges in large-scale environments, we propose a spatially aware hybrid re-initialization strategy that ensures robust reconstruction across complex scenes. Extensive experiments demonstrate that TopoSurfel achieves competitive geometric reconstruction accuracy while maintaining high-quality mesh-based novel view synthesis. The code for our method is available at https://github.com/Fan-Treasure/TopoSurfel.