TopoSurfel: Closing the Loop between Gaussian Surfels and Meshes for Surface Reconstruction

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.

• The paper introduces TopoSurfel, a methodology for enhancing 3D Gaussian Splatting surface reconstruction by integrating Gaussian surfels with continuous meshes, improving geometric fidelity and mesh extraction efficiency.
• Testing on various datasets, TopoSurfel achieves superior geometry and visualization results, often outpacing competitors in term of surface completion and fidelity, having a mean Chamfer distance of 0.51 mm on DTU.
• On TopsoSurfel's rendering efficiency and scalability, it operates using 0.14M Gaussians and 8.5GB peak memory, with optimization objectives including photometric loss ensuring higher fidelity.

TopoSurfel addresses a persistent weakness of 3D Gaussian Splatting (3DGS): although 3DGS excels at novel view synthesis, its discrete, unstructured primitives make reliable surface extraction difficult. Existing reconstruction methods such as SuGaR, 2DGS, PGSR, and QGS impose local geometric regularizers (depth, normal, multi-view consistency), but these constraints act only where observations are strong and do not propagate structural information across textureless or occluded regions. TopoSurfel's central idea is to close the optimization loop between Gaussian surfels and continuous meshes: at every training iteration, the surfel set is differentiably converted into an explicit proxy mesh that serves as a global structured prior, without introducing any auxiliary neural networks or extra per-Gaussian learnable parameters.

Motivation and relation to prior work

The paper positions itself against two families of approaches. Neural implicit methods (NeuS, VolSDF, Neuralangelo) achieve high geometric fidelity but train slowly and require post-processing for topology extraction. Gaussian-based methods improve surface fitting through primitive redesign or regularization, but typically extract meshes only after training via TSDF fusion, leaving geometry outside the optimization loop. The closest prior is MILo, which first achieved differentiable mesh extraction during 3DGS training; however, MILo operates on volumetric Gaussians and relies on virtual-corner subdivision plus DMTet, increasing computational and memory cost while keeping the Gaussian-to-mesh link indirect. TopoSurfel instead builds on planar surfel representations (2DGS, Gaussian Surfels, PGSR), which are closer to actual surfaces, and establishes a direct, parameter-free mapping from surfels to meshes.

Method

Spatially-aware hybrid warm-up

Extracting a mesh directly from sparse SfM points yields a scattered "Gaussian soup" that destabilizes differentiable iso-surfacing. TopoSurfel therefore runs a 10,000-iteration warm-up with photometric loss, scale regularization to flatten Gaussians into surfels, and PGSR-style multi-view consistency loss. A coarse TSDF mesh is then extracted, and one surfel is initialized at each triangle centroid using the face frame for rotation and tangential scales — a one-to-one discrete approximation of the mesh. For large-scale scenes, TSDF fusion often covers only the foreground object; reinitializing all surfels from this partial mesh would destroy background representation. The hybrid strategy therefore reinitializes only surfels within $\beta \cdot D_{\text{scene}}$ of the mesh and retains optimized background surfels elsewhere.

Differentiable proxy mesh extraction

The mesh branch introduces no trainable parameters and proceeds in three steps. First, opacity-filtered surfels are sampled into a weighted oriented point cloud: each surfel contributes a center point (weight $\alpha_i$) and four tangent-plane offset points at two standard deviations (weight $0.5\alpha_i$), all carrying the oriented surfel normal. Second, differentiable Poisson surface reconstruction (DPSR, following Shape as Points) solves for an indicator field in the frequency domain via FFT/IFFT on a $512^3$ voxel grid. Third, differentiable marching cubes (DiffMC) extracts vertices by linear interpolation along voxel edges straddling the iso-surface, making vertex coordinates continuous functions of the scalar field. Gradients flow back to surfel parameters through this entirely parameter-free construction.

Mesh-guided surfel evolution

Because surfels capture orientation but not front/back sides, normal orientation combines two cues: when a surfel lies within $\gamma \cdot \text{scale}_i$ of the proxy mesh and its unoriented normal agrees with the nearest face normal ($$|\mathbf{n}_i \cdot \hat{\mathbf{n}}_i| \geq \tau_{\cos}$$), the mesh face normal serves as reference; otherwise, an accumulated viewing-direction vector over visible cameras determines the flip sign. Density control is augmented with explicit point-to-face distances: low-opacity surfels far from the extracted surface are pruned as floaters, and any mesh face that is no surfel's nearest neighbor indicates a coverage gap and triggers densification at the face centroid, inheriting geometric parameters from the face and SH coefficients from the nearest existing surfel.

Optimization objectives

After warm-up, both the surfels and the proxy mesh are rendered (the latter via nvdiffrast), and log-depth and normal consistency losses between the two renderings provide bidirectional supervision, following MILo's formulation. The total objective combines photometric loss, scale regularization, multi-view consistency, and the mesh consistency terms gated by an indicator activating after warm-up.

Experimental results

On DTU, TopoSurfel achieves a mean Chamfer distance of 0.51 mm, second best overall and the best among methods without monocular depth priors, in 37 minutes — compared to PGSR's 0.53 mm (30 min), QGS's 0.54 mm (48 min), and MILo's 0.68 mm (43 min). On Tanks and Temples it attains the best mean F1-score of 0.52 in 98 minutes, versus 0.50 for Neuralangelo, PGSR, and QGS, and 0.49 for dense-configured MILo (182 min). On Mip-NeRF 360, rendering quality remains competitive (30.47 dB indoor PSNR, 24.18 dB outdoor), indicating that the geometric prior does not sacrifice visual fidelity. On NeRF-Synthetic mesh-based NVS — where fixed meshes are UV-unwrapped and only textures are optimized — TopoSurfel achieves the best results (25.23 dB PSNR, 0.921 SSIM, 0.087 LPIPS), a notable margin over GOF (23.97 dB) and MILo (23.77 dB), suggesting the extracted meshes are more complete and topologically usable than those of competing methods.

Ablations on TNT quantify each component: removing the mesh loop drops F1 from 0.52 to 0.42; removing warm-up drops it to 0.39; removing hybrid initialization causes a severe collapse to 0.21 F1 and 20.59 dB PSNR; random normal flips degrade to 0.37. Supplementary ablations show five-point sampling outperforms single-point sampling (F1 0.53 vs 0.49 with KNN-density weighting, though slower), DiffMC at $512^3$ resolution outperforms FlexiCubes constrained to $400^3$, and discarding Gaussian flattening in favor of GOF/RaDe-GS style normal estimation destabilizes the closed loop because noisy proxy meshes propagate incorrect constraints back to the surfels.

Efficiency and scalability

TopoSurfel uses 0.14M Gaussians, 8.5 GB peak memory, and trains in 37 minutes on an RTX 3090, rendering at 284 FPS with a 1.6M-vertex output mesh — a favorable balance against GOF (10.1 GB, 41 FPS) and MILo (9.4 GB, 114 FPS). A memory stress test reveals the principal bottleneck: DPSR's dense volumetric grid scales cubically with resolution. Raising the grid from $256^3$ to $512^3$ improves TNT F1 from 0.45 to 0.52, but at $\alpha_i$0 three TNT scenes exceed 24 GB of GPU memory. Consequently, the differentiable mesh serves only as a training-time structural proxy at $\alpha_i$1, while the final mesh is exported via non-differentiable TSDF fusion, since final export requires no gradient flow.

Limitations and open questions

The paper concedes several limitations explicitly. First, differentiable iso-surfacing resolution is memory-bounded, limiting recovery of high-frequency micro-geometry in unbounded scenes relative to offline post-processing; whether adaptive or sparse spatial representations can remove this cubic scaling remains open. Second, like other 3DGS-based methods, highly specular or transparent surfaces remain difficult because Gaussians model view-dependent radiance rather than underlying geometry — the Materials scene failure case illustrates incomplete surface recovery under strong view-dependent appearance. Third, the method depends on TSDF fusion quality during initialization, so scenes where depth truncation prevents even coarse foreground extraction fall outside the demonstrated regime. Finally, the proxy mesh is discarded at export rather than being the deliverable, raising the question of whether higher-resolution or hierarchical differentiable extraction could produce final meshes directly.

Conclusion

TopoSurfel demonstrates that a parameter-free, fully differentiable surfel-to-mesh conversion — DPSR followed by DiffMC — can inject a global topological prior into Gaussian surfel optimization, yielding state-of-the-art geometry on DTU and TNT at competitive cost while preserving rendering fidelity and producing meshes that outperform alternatives in downstream textured rendering. Its main constraint is the memory cost of dense volumetric iso-surfacing, which currently caps the achievable detail in large-scale scenes.