Abstract
Point-based rendering has emerged as a robust alternative to traditional triangle meshes for visualizing massive 3D datasets. However, standard point-based techniques utilizing flat "surfels" (surface elements) often suffer from jagged, unrealistic silhouettes. This paper introduces c-surfels—curved surface elements—designed to adapt to underlying geometry. By selectively applying these curved primitives to an object's silhouette while maintaining planar surfels elsewhere, the proposed hybrid pipeline achieves significantly smoother visual results with minimal computational overhead.
Introduction
As 3D scanning technology has evolved, the scale of point cloud data has grown from thousands to billions of points. Traditionally, these points are converted into triangle meshes for rendering. However, the sheer volume of data makes the overhead of maintaining connectivity—knowing which vertex connects to which—prohibitively expensive.
Point-based rendering bypasses the mesh entirely, treating each point as an independent disk called a surfel. While efficient, standard surfels are planar. When viewed at the edge of an object—the silhouette—these flat disks overlap to create a "blocky" or "stepped" appearance. This research addresses the challenge of improving visual quality at these edges without sacrificing the speed and simplicity that make point-based rendering attractive.
Methodology
The authors propose a hybrid rendering pipeline that combines standard planar surfels with a new primitive: the c-surfel.
1. The C-Surfel Primitive
Unlike a flat disk, the c-surfel is a curved patch defined by a mathematical function ($\phi(r)$) that approximates the local surface. It acts as a small, deformed piece of a plane that follows the geometric "flow" of the object. By treating each curved patch as a homeomorphism (a deformed piece of the plane), the authors can apply standard 2D texture mapping techniques to 3D curved surfaces without complex re-parameterization.
2. Selective Silhouette Rendering
To maintain high performance, the system does not apply curvature to every point. Instead, a "silhouette test" is performed by calculating the angle between the observer’s view vector and the point’s normal. If the angle is near 90 degrees—indicating the point lies on the edge—a c-surfel is used. Otherwise, the system defaults to a standard planar surfel. This "silhouette priority" leverages the human eye's sensitivity to edge errors, focusing computational effort where it is most visually impactful.
3. Geometry Analysis and Blending
To ensure surface continuity, the authors utilize Principal Component Analysis (PCA) on local point neighborhoods to estimate surface normals and appropriate surfel sizes. For rendering, a Gaussian alpha mask is applied to each surfel, creating transparency that fades toward the edges. Overlapping surfels are then blended using a "normalized weighted mean," ensuring the final color is a smooth average of the contributing points rather than a collection of discrete disks.
Experimental Results
The method was evaluated using several high-density point cloud models, including the Igea (134,345 points), Ball Joint (137,059 points), Buddha (389,347 points), and Rabbit (44,691 points).
Qualitative and Quantitative Performance
The visual difference at the silhouette is striking: while planar surfels create a "sawtooth" edge, c-surfels produce a smooth, rounded contour comparable to high-resolution meshes.
Quantitatively, the hybrid approach proved highly efficient. Benchmarks performed on the Igea model (using a Pentium IV 1.4 GHz system) demonstrated the following:
- Purely Planar Rendering: 1.22 FPS
- Purely Curved Rendering: 1.06 FPS
- Hybrid (Planar + Silhouette c-surfels): 1.18 FPS
The hybrid method provides nearly the same visual quality as the fully curved version while running at approximately 96% of the speed of the basic planar version.
Discussion and Future Work
The research reinforces the idea that if point density is sufficiently high, explicit connectivity is unnecessary. Smart alpha-blending and adaptive primitive shapes can effectively simulate a solid surface.
However, limitations remain. High-curvature regions with sharp spikes or complex local folds still pose challenges for the Gaussian approximation used in c-surfels. Furthermore, the original implementation was written in unoptimized C; the authors suggest that migrating these calculations to programmable GPUs would drastically increase frame rates. Future work also includes utilizing these reconstructed normals for advanced image-based effects, such as realistic relighting of scanned objects in virtual environments.
Summary updated