Fast Synthesis and Rendering of BTF on Arbitrary Surfaces

Zhan Zhang, Yue Qi, Yong Hu2008 · DOI

Abstract

This paper proposes an approach for BTF (Bidirectional Texture Function) synthesis and rendering. A BTF can be regarded as a kind of multi-dimensional texture, which preciously shows self-shadow, self-occlusion and inter-reflection appearance of object surfaces under varying view-and light-directions. However, it is difficult to use BTF on object surfaces because of its huge data and small size. Our method uses PCA algorithm to compress BTF data, and then we propose a Wang tiles based synthesis algorithm to solve its size problem. To keep consistency of multi-dimensional, we also propose an error determination algorithm for Wang tiles' construction in synthesis process. Using the method we proposed can render BTF on arbitrary object surfaces efficiently and quickly.

TL;DR

This paper presents a streamlined pipeline for applying Bidirectional Texture Functions (BTFs)—high-fidelity textures that capture complex light effects like self-shadowing—to arbitrary 3D surfaces. By combining PCA-based compression with a Wang tiles synthesis strategy, the authors achieve realistic, real-time rendering of complex materials at over 100 frames per second.

What Problem Does This Paper Address?

In computer graphics, standard 2D textures often look "flat" because they don't account for how a surface’s appearance changes based on the light source or the viewer's perspective. Bidirectional Texture Functions (BTFs) solve this by capturing 6D data: a set of images of a material taken from thousands of different light and view combinations. This allows for incredibly realistic effects like self-occlusion, inter-reflection, and self-shadowing.

However, BTFs face two major hurdles:

  1. Data Volume: A single BTF sample can consist of thousands of high-resolution images, making it too large for standard memory.
  2. Sample Size: Real-world BTF samples are usually captured as small, flat squares. Mapping these onto a large, complex 3D model (like a character or a vehicle) requires "synthesizing" the texture to cover the surface without showing obvious seams or repetitive patterns.

Methods

The authors propose a three-stage pipeline to make BTFs practical for real-time applications:

The proposed BTF synthesis and rendering pipeline.
  • Compression via PCA: They use Principal Component Analysis (PCA) to reduce the massive image set. By extracting the most important "eigen-textures," they can reconstruct the original appearance using only a few components (typically around 5), significantly reducing the memory footprint.
  • Synthesis via Wang Tiles: To cover large surfaces, the authors use Wang tiles—a set of tiles designed so their edges match up perfectly. They propose a synthesis algorithm that chooses four random patches from the compressed BTF and uses a "cutting path" (calculated via Dijkstra’s algorithm) to create seamless transitions.
Illustration of the cutting path computation for Wang tile construction.
  • Multi-Dimensional Error Computation: A key novelty is their error determination algorithm. When synthesizing tiles, they don't just look at one image; they calculate the error across all light and view directions simultaneously using the PCA coefficients. This ensures the texture remains consistent even as the lighting moves.
  • GPU Rendering: The final rendering is handled by GLSL shaders. The PCA weights are stored in 2D textures, and the shader performs tri-linear interpolation between the nearest captured light/view directions to produce the final pixel color on the fly.

Datasets

The researchers evaluated their method using the Bonn BTF database, specifically focusing on two material samples: WOOL (a highly textured fabric) and IMPALLA (a complex, porous material). Each dataset contained 6,561 sample images (81 view directions × 81 light directions). For the 3D geometry, they used the standard Stanford Bunny (30,660 triangles) and a Shark model (20,588 triangles).

Key Insights and Ideas

  • BTF as a "c-channel" Image: The authors treat the compressed BTF as having c channels (where c is the number of PCA components). This allows them to treat high-dimensional light data as a standard image synthesis problem, making it computationally tractable.
  • Consistency is King: The "aha moment" in their synthesis is the error thresholding. By ensuring the error $d$ is less than a calculated $d_{max}$ across the entire light/view manifold, they prevent visual artifacts that usually occur when a texture looks good from one angle but breaks from another.
  • Efficiency vs. Quality: They demonstrated that using only 5 principal components results in a reconstruction error of less than 2.22%, proving that the vast majority of BTF data is redundant for human perception.

Key Results

  • Compression: They achieved a compression ratio of 6.17%, making the data small enough to fit into GPU memory.
  • Speed: The system achieved impressive frame rates: 106 FPS for the bunny model and 134 FPS for the shark.
  • Visual Fidelity: The reconstructed images maintained the complex self-shadowing and highlights of the original raw data, even when mapped onto non-planar, arbitrary 3D surfaces.

Limitations and Future Work

The authors acknowledge a few key limitations. The current synthesis method works well for isotropic materials but struggles with anisotropic patterns (like brushed metal) because the edge-matching logic doesn't currently account for directional orientation. Future work includes developing methods to blend multiple different BTF materials on a single mesh and further optimizing the pre-computation time, which currently takes between 32 and 44 minutes.

Summary updated

Abstract

This paper proposes an approach for BTF (Bidirectional Texture Function) synthesis and rendering. A BTF can be regarded as a kind of multi-dimensional texture, which preciously shows self-shadow, self-occlusion and inter-reflection appearance of object surfaces under varying view-and light-directions. However, it is difficult to use BTF on object surfaces because of its huge data and small size. Our method uses PCA algorithm to compress BTF data, and then we propose a Wang tiles based synthesis algorithm to solve its size problem. To keep consistency of multi-dimensional, we also propose an error determination algorithm for Wang tiles' construction in synthesis process. Using the method we proposed can render BTF on arbitrary object surfaces efficiently and quickly.
BibTeX
@misc{zhang2008fast,
  author = {Zhan Zhang and Yue Qi and Yong Hu},
  title = {{Fast Synthesis and Rendering of BTF on Arbitrary Surfaces}},
  year = {2008},
  doi = {10.2991/jcis.2008.63},
}
Figure