Acquiring Bidirectional Texture Functions for Large-Scale Material Samples

Heinz Christian Steinhausen, Dennis den Brok, Matthias B. Hullin, Reinhard KleinJ. WSCG 2014

Introduction

Bidirectional Texture Functions (BTFs) represent the pinnacle of material realism in computer graphics. Unlike simple textures, BTFs capture how light interacts with complex surfaces from various angles, accounting for phenomena like self-shadowing, inter-reflections, and parallax. However, a significant "size bottleneck" has long hindered their practical application: high-end acquisition devices, such as camera domes, are typically limited to small samples (around 10cm x 10cm).

When dealing with materials that exhibit large-scale structures—such as the sweeping grain of a wooden plank or intricate wallpaper patterns—simply tiling a small captured patch leads to repetitive, unrealistic results. Capturing these materials at full scale is often impossible due to hardware constraints and the massive data volumes involved. This paper presents a novel hybrid approach to bridge this gap, enabling the reconstruction of high-quality, large-scale BTFs from limited measurements.

Methodology: A Hybrid Synthesis Pipeline

The authors propose a three-step pipeline designed to synthesize large-scale material appearance while maintaining the physical accuracy of dense measurements.

1. Representative Measurement

The process begins with a full, dense BTF measurement of a small spatial region. This patch must contain the essential "ingredients" of the material—its fundamental reflectance properties and local geometric structures.

2. Large-Scale Constraints

To guide the synthesis over a larger area, the system utilizes sparse "constraint" images of the full-scale sample. Remarkably, the authors found that even a single "mean color image"—an average of the material under various lighting and viewing conditions—can effectively map out large-scale structures like wood knots or grain patterns.

3. Controlled Texture Synthesis in the SVD Domain

The core of the reconstruction is an adapted "Texture Optimization" algorithm. To manage the high dimensionality of BTF data, the synthesis is performed directly on a compressed representation using Singular Value Decomposition (SVD). Instead of raw pixels, the algorithm operates on "eigen-ABRDFs," significantly reducing the computational and memory footprint.

To combat the blurring often associated with optimization-based synthesis, the pipeline includes an iterative refinement step. This step alternates between satisfying global constraints and restoring high-frequency details through statistical synthesis, ensuring the final result retains the "feel" and sharpness of the original material.

Evaluation and Key Insights

The researchers evaluated their method using custom-captured data from a specialized "Dome I" device featuring 151 cameras and light sources. Testing on diverse materials including carpet, leather, and wood revealed several key insights:

  • BTF Redundancy: The success of the method relies on the fact that BTF data is highly redundant; different parts of a material often share the same underlying reflectance properties in different spatial arrangements.
  • Constraint Efficiency: Using a mean color image as a constraint is a highly effective "hack," allowing for the capture of macro-structures without requiring high-end equipment for the full-scale sample.
  • Visual and Quantitative Success: Synthesized results for wood and leather were often indistinguishable from ground truth. Quantitative analysis using Root Mean Square Error (RMSE) confirmed high accuracy, even when using only a single constraint image.

Limitations and Future Directions

While the approach significantly improves the scalability of BTF acquisition, challenges remain. The synthesis process is computationally intensive, with a 512x512 patch taking approximately one hour to process. Future work may involve moving toward faster linear model fitting to reduce this overhead. Additionally, aligning "in the wild" handheld photos of large objects with lab-captured patches presents a calibration challenge that must be addressed for broader real-world application.

By combining dense local measurements with sparse global constraints, this research provides a practical path forward for digitizing the complex, large-scale materials that define our physical world.

Summary updated

BibTeX
@article{steinhausen2014acquiring,
  author = {Heinz Christian Steinhausen and Dennis den Brok and Matthias B. Hullin and Reinhard Klein},
  title = {{Acquiring Bidirectional Texture Functions for Large-Scale Material Samples}},
  journal = {J. WSCG},
  year = {2014},
}
Figure