GSPlane: Structured Planar Reconstruction

• GSPlane is a structured method that enhances planar reconstruction by reparameterizing Gaussian centers using convex combinations of plane basis points.
• It employs a dynamic Gaussian re-classifier to mitigate misclassification errors, ensuring optimal convergence and mesh accuracy during training.
• The approach facilitates editable scene manipulation by producing topologically coherent, compact meshes with improved geometric fidelity.

GSPlane is a structured approach for concise and accurate 3D planar reconstruction within Gaussian Splatting frameworks, specifically developed to address the challenges of geometrically consistent plane extraction and surface mesh refinement in scenes dominated by planar regions. By integrating robust planar priors derived from 2D segmentation and normal estimation, GSPlane enforces geometric consistency during optimization and yields well-structured, topologically correct mesh representations. The methodology includes dynamic strategies to ensure classification reliability and offers direct utility for scene editing and downstream simulation tasks, with empirical evidence of substantial improvements in geometric fidelity and mesh compactness.

1. Motivation and Background

Gaussian Splatting (GS) has become a prominent paradigm for novel view synthesis and explicit surface reconstruction in 3D vision, providing photorealistic renderings by representing scenes as collections of Gaussian primitives. Typical environments—especially indoor and urban scenes—are rich with planar surfaces, which present obstacles for vanilla GS approaches due to inadequate smoothness and over-dense mesh connectivity in reconstructed plane regions. GSPlane (Gan et al., 20 Oct 2025) was introduced to reconstruct planar geometry with a structured and parameterized representation, overcoming the limitations of previous GS methods and facilitating downstream physical simulation, editing, and digital twin generation.

2. Structured Representation via Planar Priors

GSPlane’s fundamental innovation is its use of structured planar priors to reparameterize Gaussian coordinates for planar regions. The approach begins with the extraction of planar masks and normals from each posed image, utilizing established off-the-shelf models—SAM for segmentation masks and Metric3Dv2 for per-pixel surface normals. For each mask, the average mask normal is calculated, and pixels whose normal cosine similarity exceeds a threshold are identified. Regions with more than 70% of pixels meeting this criterion are classified as planar.

These 2D planar masks are projected into 3D via an initial point cloud or sparse reconstruction. The Gaussian centers associated with each detected planar region are then parameterized not as Cartesian coordinates but as convex combinations of three non-collinear basis points on the plane, enforcing collinearity and planarity:

$$V'_p = \omega_1 F_1 + \omega_2 F_2 + \omega_3 F_3 \quad \text{where} \quad \omega_1 + \omega_2 + \omega_3 = 1$$

Both the basis points $\{F_1, F_2, F_3\}$ and weights $\{\omega_i\}$ are optimized during training, ensuring that assigned Gaussians adhere strictly to the inferred plane geometry.

3. Training Robustness: Dynamic Gaussian Re-classifier

To mitigate errors from mask or normal estimation (e.g., false positives in planar region assignment), GSPlane employs a Dynamic Gaussian Re-classifier (DGR). During optimization, the system tracks the gradient magnitude for every Gaussian. If a Gaussian assigned to a planar cluster persistently exhibits high gradients—specifically, exceeding the average of the top 20% gradients among all non-planar Gaussians—it is reclassified as non-planar. This entails reverting its structured parameterization back to the original Cartesian formulation. The DGR process is performed iteratively throughout training, notably during densification phases, thereby increasing the reliability and convergence of the structured representation.

4. Mesh Layout Refinement via Planar Priors

Mesh refinement leverages optimized planar priors to address density and topological inconsistency in post-reconstructions. After training, mesh vertices are grouped into planar clusters by finding proximity to the original point cloud (thresholded by voxel size $\delta$). Mesh faces entirely composed of vertices on the same plane are identified, and their vertices are categorized into boundary and interior sets.

Interior vertices are projected onto 2D grids using minimum enclosing rectangle (MER) alignment to accommodate irregular shapes. Delaunay triangulation is applied to combine grid points and boundary vertices, followed by remapping back to 3D using the optimized plane basis. This procedure yields meshes with significantly fewer vertices, unified normal consistency, and topological correction compared to standard GS outputs.

5. Applications: Decoupled Scene Editing and Supportive Plane Manipulation

GSPlane’s structured planar representation enables advanced scene manipulation. In Supportive Plane Correction (SPC), planar regions (e.g., floors, tables) are isolated, allowing objects residing atop these planes to be decoupled and flexibly repositioned. With sealed contact boundaries and clean planar geometry, objects can be removed or moved without affecting surface integrity—beneficial for digital twin modeling, physics simulation, and interactive editing.

6. Experimental Results: Accuracy, Compactness, and Rendering Fidelity

Empirical validation on datasets such as ScanNetV2 (indoor) and Tanks and Temples (outdoor) demonstrates GSPlane’s superiority. Quantitative metrics show improved geometric accuracy—measured by Accuracy, Completion, F-score—over baseline GS and plane extraction methods. Meshes produced via GSPlane maintain reduced vertex and face counts, yet preserve rendering quality (PSNR, SSIM, LPIPS). Planar-wise evaluations report lower Chamfer distances for the largest planes, highlighting the method’s efficacy in structural reconstruction. Inclusion of the DGR and parameterized representation is shown to outperform off-the-shelf supervision approaches, underscoring the importance of these techniques for robust optimization.

7. Limitations and Research Directions

GSPlane’s current focus is restricted to planar regions, while structured representation for non-planar surfaces remains unresolved. Future enhancements may include improved segmentation/normal predictors and extension of the structured modeling paradigm to curved or irregular surfaces. A plausible implication is that further robustness in planar prior extraction could catalyze more accurate digital twin pipelines and interactive scene editors for complex environments.

---

GSPlane establishes a rigorous foundation for concise and accurate planar reconstruction within Gaussian Splatting frameworks, combining geometric priors, dynamic adaptive classification, and mesh refinement procedures to deliver topologically coherent, editable, and simulation-ready planar scene models.