GPU Rendering of Relief Mapped Conical Frusta

D. Bhagvat, Stefan Jeschke, David Cline, Peter WonkaComputer Graphics Forum 2009 · DOI

TL;DR

This paper introduces a novel method for rendering complex skeletal objects by representing them as a series of relief-mapped conical frusta. This approach leverages GPU ray casting to achieve high-quality, interactive rendering with significantly reduced CPU workload and data transfer, making it ideal for complex scenes and animations.

What Problem Does This Paper Address?

Rendering complex 3D objects, especially those with intricate details and animations, often faces bottlenecks due to data bandwidth limitations between the CPU and GPU, and the sheer memory requirements of detailed polygonal models. Traditional methods might approximate complex shapes with a vast number of triangles, leading to high memory usage and slow rendering. Alternatively, using high-level primitives can simplify data transfer but might sacrifice fine-scale detail. This paper aims to bridge this gap by enabling high-quality, detailed rendering of skeletal objects using a compact, GPU-centric representation that minimizes CPU-GPU communication and memory overhead.

Methods

The core of the paper is the representation of skeletal objects as a chain of conical frusta. Each segment of the object is modeled as a conical frustum, and these are seamlessly connected at joints using a carefully computed cutting plane.

Key techniques include:

  • Conical Frusta Representation: A mathematical definition for conical frusta is provided, including a consistent surface parametrization and a continuous surface normal definition. This ensures smooth shading and proper texture mapping.
  • Seamless Joint Connections: A method is described to connect adjacent frusta by defining a medial plane and then constructing a cutting plane that ensures a smooth transition between segments. While an iterative fitting step is mentioned, it's noted as often unnecessary in practice.
  • Frustum Parametrization: A (u, v) parametrization is defined for each frustum segment, allowing for texture and relief mapping. The u-coordinate represents the angle around the segment's axis, and the v-coordinate spans along the segment's length. Adjustments are made to align parameters across connected segments.
  • Surface Normals: A C⁰ continuous surface normal is computed by interpolating between normals at the segment ends. This provides visually smooth shading, even though it deviates slightly from the exact surface normals.
  • GPU Ray Casting: The primary rendering technique is analytical ray casting performed directly on the GPU. This allows for efficient intersection tests with the implicit conical frusta.
  • Relief Mapping: To add fine-scale surface details, the conical frusta representation is extended with relief mapping. This involves defining an inner and outer frustum to create a "shell" where a height map is applied. The GPU ray caster then marches through this shell to find intersections with the relief-mapped surface.
  • Bounding Primitives: For efficient GPU rasterization and ray casting initiation, each conical frustum is approximated by a low-polygon count bounding primitive (a pyramidal frustum).

Datasets

Not discussed in the paper. The paper focuses on a novel representation and rendering technique, using synthetic scenes and models for evaluation rather than pre-existing datasets.

Key Insights and Ideas

  • GPU Ray Casting for Implicit Surfaces is Powerful: The paper demonstrates that GPU ray casting is not just for simple primitives like spheres or quadratic surfaces, but can be effectively applied to more complex implicit shapes like conical frusta, especially when combined with relief mapping.
  • Relief Mapping on Implicit Surfaces Solves Detail Issues: Relief mapping, traditionally applied to polygonal meshes, is extended to implicit surfaces. This approach avoids common issues with polygonal relief mapping, such as silhouette artifacts and the need for excessive polygon subdivision.
  • Compact Representation for Complex Objects: Representing objects as a series of conical frusta is significantly more memory-efficient than using dense polygonal meshes, especially when relief mapping is involved. This is crucial for handling complex scenes and animations.
  • Minimal CPU-GPU Communication: By performing most of the geometric processing and rendering on the GPU, the method drastically reduces the data that needs to be transferred from the CPU, leading to higher frame rates and interactivity.
  • Smooth Shading with Interpolated Normals: The technique of interpolating surface normals across segment boundaries provides visually pleasing smooth shading, even if the normals aren't perfectly accurate to the underlying geometry.

Key Results

  • High-Quality Rendering: The method produces visually appealing renderings with fine-scale details comparable to highly subdivided polygonal models.
  • Interactive Frame Rates: The paper demonstrates interactive frame rates for both static and animated scenes, even with relief mapping applied. For example, a scene requiring at least 256k polygons for similar quality with polygonal rendering could be rendered with ray casting using far fewer implicit segments.
  • Performance Comparison: Table 1 shows that ray casting (RC) without relief mapping is generally faster than ray casting with relief mapping (RC+R). However, both are significantly faster than polygon rendering (GS) when high subdivision levels are required to match the quality of relief-mapped ray casting. For instance, the "Pine tree" model achieves 34 fps with RC and 16 fps with RC+R, while polygon rendering requires 55/1 fps (subdivision level 1) and 15/2 fps (subdivision level 2) to achieve comparable quality.
  • Compactness: The conical frustum representation is shown to be significantly more memory-efficient than polygonal models.

Limitations and Future Work

  • Piecewise Linear Segments: Objects are inherently composed of piecewise linear segments. This makes representing truly smooth, curved objects in the medial-axis direction inefficient.
  • Joint Intersections: The current method does not handle cases where the top and bottom caps of adjacent segments might intersect if the angle between them is too large.
  • Branching Structures: Seamlessly representing branching structures is not directly supported.
  • Round Cones: The basic representation assumes round cones. While relief mapping can add shape variety, extending to other implicit surfaces like ellipsoids could broaden applicability.

Future work suggested includes:

  • Providing C¹ continuity between adjacent segments for smoother object representations.
  • Developing methods to handle branching structures more effectively.
  • Exploring the use of relief maps to improve continuity in branching structures.
  • Extending the implicit surface primitives beyond conical frusta to include shapes like ellipsoids.

Summary updated

BibTeX
@article{bhagvat2009gpu,
  author = {D. Bhagvat and Stefan Jeschke and David Cline and Peter Wonka},
  title = {{GPU Rendering of Relief Mapped Conical Frusta}},
  journal = {Computer Graphics Forum},
  year = {2009},
  doi = {10.1111/j.1467-8659.2009.01420.x},
}
Figure