Topology-Aware 3D Gaussian Splatting: Leveraging Persistent Homology for Optimized Structural Integrity

Tianqi Shen, Shaohua Liu, Jiaqi Feng, Ziye Ma, Ning AnAAAI 2025 · DOI

Introduction

3D Gaussian Splatting (3D-GS) has emerged as a state-of-the-art technique for real-time 3D reconstruction and novel-view synthesis. By representing scenes as a collection of anisotropic Gaussians, it achieves high-fidelity rendering at impressive speeds. However, the quality of 3D-GS is fundamentally constrained by its initialization, which typically relies on sparse point clouds generated via Structure-from-Motion (SfM). This dependency often results in structural artifacts, particularly in regions with low texture or complex topology. This paper introduces Topology-GS, a novel framework that integrates topological data analysis (TDA) to preserve structural integrity and improve perceptual rendering quality.

The Challenge: Geometric and Topological Gaps

The reliance on SfM leads to two primary issues in standard 3D-GS optimization:

  1. Incomplete Geometric Coverage: SfM frequently fails to capture points on low-curvature or uniform surfaces (e.g., white walls, smooth tabletops). The standard Adaptive Density Control (ADC) mechanism often fails to fill these gaps meaningfully, leading to "pixel-level" structural breaks and holes.
  2. Inadequate Feature-Level Integrity: Conventional loss functions, such as L1 and Structural Similarity Index (SSIM), operate on local pixel values. They lack a global understanding of the scene's topology—the loops, voids, and connectivity of components. Consequently, while pixel values may be numerically close, the resulting images often exhibit high Learned Perceptual Image Patch Similarity (LPIPS) scores, indicating they appear "unnatural" to human observers.
Overview of the Topology-GS framework compared to baseline methods. (a) Baseline struggles with incomplete coverage and topological constraints. (b) Ground truth. (c) Topology-GS utilizes LPVI and PersLoss to achieve superior structural fidelity.

Methodology: Topology-Aware Optimization

Topology-GS addresses these limitations through two core innovations: Local Persistent Voronoi Interpolation (LPVI) and a Persistent Homology-based loss function (PersLoss).

Local Persistent Voronoi Interpolation (LPVI)

To address geometric gaps, LPVI replaces the standard random point densification with a topology-guided interpolation strategy. The process uses Voronoi tessellation to identify candidate regions for new Gaussians. Before a point is added, the system performs a "TopoDiff" check using Persistent Homology (PH)—a mathematical tool that tracks topological features across different spatial scales.

If the addition of a 3D point significantly alters the local topology (measured via Wasserstein distance), the system adaptively switches to 2D interpolation using Principal Component Analysis (PCA) projection. This ensures that flat surfaces remain planar and that the optimization does not erroneously fill intentional structural voids.

Persistent Homology Loss (PersLoss)

To enforce structural integrity during rendering, the authors propose PersLoss. This loss function treats the 2D rendered image as a 3D topological space, using RGB intensities as the third dimension. By generating "persistent barcodes"—visual summaries of topological features—for both the rendered and ground-truth images, the model can minimize the distance between their underlying structures.

Detail of the densified point cloud in Topology-GS, showing how Gaussians are organized to maintain topological features like loops and connected components.

By focusing on the "top-k" longest-lived features in these barcodes, PersLoss ignores high-frequency noise and forces the Gaussians to align with the significant structural components of the scene. This global regularizer compensates for the limitations of pixel-wise loss functions.

Experimental Evaluation

The method was rigorously evaluated on standard benchmarks, including Mip-NeRF360, Tanks & Temples, and Deep Blending.

  • Quantitative Results: Topology-GS achieved state-of-the-art performance, consistently outperforming baselines like Scaffold-GS. Notably, it showed significant improvements in LPIPS, reflecting higher perceptual quality. On the Mip-NeRF360 dataset, it reached a PSNR of 29.50, compared to the baseline's 28.84.
  • Qualitative Improvements: Rendered depth maps demonstrated smoother, more continuous surfaces in previously problematic flat regions. Feature map visualizations (via EigenCAM) confirmed that the model's internal representations more closely mirror the structural patterns of the ground truth.
  • Efficiency: Despite the complexity of PH calculations, the authors optimized the process by truncating the analysis to significant features and only activating the topological constraints during specific ADC stages, resulting in a negligible memory overhead (~20MB).

Conclusion

Topology-GS demonstrates that integrating topological priors into the 3D Gaussian Splatting pipeline effectively resolves long-standing issues with geometric sparsity and structural inconsistency. By leveraging Persistent Homology as both an interpolation guide and a regularization term, the framework produces reconstructions that are not only numerically accurate but also perceptually faithful to the real-world environment. Future work may extend these topological constraints to dynamic 4D scenes and more compact scene representations.

Summary updated

BibTeX
@inproceedings{shen2025topology,
  author = {Tianqi Shen and Shaohua Liu and Jiaqi Feng and Ziye Ma and Ning An},
  title = {{Topology-Aware 3D Gaussian Splatting: Leveraging Persistent Homology for Optimized Structural Integrity}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i7.32732},
}
Figure