SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction

Xinran Yang, Donghao Ji, Yuanqi Li, Jie Guo, Yanwen Guo, Junyuan Xie

Abstract

Neural rendering techniques have made substantial progress in generating photo-realistic 3D scenes. The latest 3D Gaussian Splatting (3DGS) technique has achieved high-quality novel view synthesis and fast rendering speeds. However, 3D Gaussians lack proficiency in defining accurate 3D geometric structures despite their explicit primitive representations. This is because Gaussian attributes are primarily tailored and fine-tuned for rendering 2D images via their anisotropic nature. To pave the way for efficient 3D reconstruction, we present Spherical Gaussians, a simple and effective representation for 3D geometric boundaries. From these, we can directly reconstruct 3D feature curves from a set of calibrated multi-view images. Spherical Gaussians are optimized from grid initialization with a view-based rendering loss, where a 2D edge map is rendered at a specific view and compared to the ground-truth edge map extracted from the corresponding image, without the need for 3D supervision. We further introduce a novel optimization-based algorithm called SGCR to extract accurate parametric curves from aligned Spherical Gaussians. We demonstrate that SGCR outperforms existing state-of-the-art methods in 3D edge reconstruction while enjoying significant efficiency gains.

Introduction

While 3D Gaussian Splatting (3DGS) has revolutionized photorealistic novel view synthesis, it is notoriously poor at capturing precise geometry. The "splats" in standard 3DGS are anisotropic—stretched ellipsoids optimized solely to minimize pixel-wise color error. This often results in a "messy" internal structure that fails to align with actual surfaces or structural edges.

Reconstructing 3D feature curves—the structural "skeleton" of an object—is critical for CAD modeling, robotics, and scene understanding. Existing methods for this task typically require dense 3D point clouds or rely on coordinate-based neural fields (such as NEF or EMAP) that are computationally expensive and slow to train. SGCR addresses this by bridging the gap between 2D edge detections and a coherent 3D parametric representation without requiring 3D supervision, leveraging the speed of Gaussian rasterization.

Methodology

The SGCR pipeline consists of two primary stages: the generation of Spherical Gaussians and the extraction of parametric curves.

1. Spherical Gaussian Generation

Instead of standard 3DGS, the method utilizes Spherical Gaussians (SG). By constraining the Gaussians to be isotropic, the authors remove rotation and scaling attributes, forcing every Gaussian to be a sphere with a fixed radius. This constraint makes them act like "atoms" that naturally align along 1D structures rather than flattening into 2D surfaces.

Key components of this stage include:

  • Edge-Aware Loss: The standard color loss is replaced with a specialized loss function that balances the impact of sparse edge pixels and empty space in 2D edge maps.
  • Opacity-Color Loss: To handle occlusions—where an edge is visible in one view but blocked in another—this loss keeps the Gaussian's opacity and color intensity consistent. This prevents the optimizer from prematurely pruning valid 3D edges that are only visible from specific angles.

2. Parametric Curve Extraction (SGCR)

Once the SGs are clustered along the object's edges, the SGCR algorithm extracts smooth, usable curves through a two-step process:

  • Line Fitting: A coarse stage uses a randomized search and Chamfer distance to fit straight line segments to the centers of the optimized SGs.
  • Global Optimization: These segments are refined into 3rd-order rational Bézier curves. Rational Béziers are specifically chosen because they can perfectly represent circular arcs—common in manufactured objects—which standard Bézier curves cannot fit accurately.

Experimental Results

The method was evaluated on four diverse datasets: ABC-NEF (CAD models), ModelNet (furniture), and DTU & Replica (real-world indoor and outdoor scenes).

Performance and Accuracy

On the ABC-NEF dataset, SGCR achieved a Chamfer Distance (CD) of 0.0280, outperforming previous leaders like EMAP (0.0291) and NEF (0.0353). Qualitatively, SGCR maintains better connectivity in thin structures and handles sharp corners more faithfully than implicit methods, which often produce fragmented or chaotic line segments.

Efficiency

By leveraging the high-speed rasterization pipeline of 3DGS, SGCR avoids the "query-heavy" nature of implicit neural fields. The entire training process for a single scene takes approximately one minute. This represents a 50x speedup over NEF and EMAP, which typically require hours of training per scene.

Discussion and Insights

  • Constraints as a Catalyst: Counter-intuitively, removing degrees of freedom from 3DGS (fixing the shape to a sphere) makes it a much better geometric primitive for structural tasks. Isotropy forces the Gaussians to distribute themselves logically along 1D features.
  • The "Bridge" Concept: Spherical Gaussians serve as a robust intermediate representation. They are discrete enough to be optimized via 2D rendering but structured enough to allow for direct parametric curve fitting.

Limitations

The quality of the 3D reconstruction remains bottlenecked by the accuracy of the 2D edge detector (e.g., PiDiNet). If the 2D detector misses a faint edge, the 3D model will likely contain a gap. Additionally, while a fixed radius ($r_0 = 0.005$) proved effective for the tested datasets, extremely large or small objects might require hyperparameter tuning.

Conclusion

SGCR demonstrates that 3D Gaussian Splatting can be effectively repurposed for structural geometry reconstruction. By simplifying the Gaussian primitive and focusing on edge-based rendering losses, the method achieves state-of-the-art accuracy in 3D curve extraction while offering a massive leap in computational efficiency. Future work may extend this "geometry-first" approach to full surface reconstruction or dynamic scene analysis.

Summary updated

Abstract

Neural rendering techniques have made substantial progress in generating photo-realistic 3D scenes. The latest 3D Gaussian Splatting technique has achieved high quality novel view synthesis as well as fast rendering speed. However, 3D Gaussians lack proficiency in defining accurate 3D geometric structures despite their explicit primitive representations. This is due to the fact that Gaussian's attributes are primarily tailored and fine-tuned for rendering diverse 2D images by their anisotropic nature. To pave the way for efficient 3D reconstruction, we present Spherical Gaussians, a simple and effective representation for 3D geometric boundaries, from which we can directly reconstruct 3D feature curves from a set of calibrated multi-view images. Spherical Gaussians is optimized from grid initialization with a view-based rendering loss, where a 2D edge map is rendered at a specific view and then compared to the ground-truth edge map extracted from the corresponding image, without the need for any 3D guidance or supervision. Given Spherical Gaussians serve as intermedia for the robust edge representation, we further introduce a novel optimization-based algorithm called SGCR to directly extract accurate parametric curves from aligned Spherical Gaussians. We demonstrate that SGCR outperforms existing state-of-the-art methods in 3D edge reconstruction while enjoying great efficiency.
BibTeX
@inproceedings{yang2025sgcr,
  author = {Xinran Yang and Donghao Ji and Yuanqi Li and Jie Guo and Yanwen Guo and Junyuan Xie},
  title = {{SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.00544},
}
Figure