Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning to Learn and Sample BRDFs

Chen Liu, Michael Fischer, Tobias Ritschel

TL;DR

This work tackles the bottleneck of acquiring and learning neural BRDFs by jointly optimizing the BRDF model and the physical sampling pattern through meta-learning. It extends MAML-based meta-learning to also optimize the sampling coordinates (meta-sampling), using a barrier function to keep samples in valid angular regions and decoupled inner/outer loops for stability. Evaluations across Phong, Cook-Torrance, Linear, and Neural BRDFs on MERL and additional datasets show that the approach achieves up to five orders of magnitude fewer acquisition samples with comparable or better quality, and that the learned sampling patterns generalize to unseen BRDFs. The findings suggest substantial practical impact for rapid, high-fidelity BRDF acquisition and open avenues for applying meta-sampling to other nonlinear, sample-intensive domains.

Abstract

We propose a method to accelerate the joint process of physically acquiring and learning neural Bi-directional Reflectance Distribution Function (BRDF) models. While BRDF learning alone can be accelerated by meta-learning, acquisition remains slow as it relies on a mechanical process. We show that meta-learning can be extended to optimize the physical sampling pattern, too. After our method has been meta-trained for a set of fully-sampled BRDFs, it is able to quickly train on new BRDFs with up to five orders of magnitude fewer physical acquisition samples at similar quality. Our approach also extends to other linear and non-linear BRDF models, which we show in an extensive evaluation.

Learning to Learn and Sample BRDFs

TL;DR

This work tackles the bottleneck of acquiring and learning neural BRDFs by jointly optimizing the BRDF model and the physical sampling pattern through meta-learning. It extends MAML-based meta-learning to also optimize the sampling coordinates (meta-sampling), using a barrier function to keep samples in valid angular regions and decoupled inner/outer loops for stability. Evaluations across Phong, Cook-Torrance, Linear, and Neural BRDFs on MERL and additional datasets show that the approach achieves up to five orders of magnitude fewer acquisition samples with comparable or better quality, and that the learned sampling patterns generalize to unseen BRDFs. The findings suggest substantial practical impact for rapid, high-fidelity BRDF acquisition and open avenues for applying meta-sampling to other nonlinear, sample-intensive domains.

Abstract

We propose a method to accelerate the joint process of physically acquiring and learning neural Bi-directional Reflectance Distribution Function (BRDF) models. While BRDF learning alone can be accelerated by meta-learning, acquisition remains slow as it relies on a mechanical process. We show that meta-learning can be extended to optimize the physical sampling pattern, too. After our method has been meta-trained for a set of fully-sampled BRDFs, it is able to quickly train on new BRDFs with up to five orders of magnitude fewer physical acquisition samples at similar quality. Our approach also extends to other linear and non-linear BRDF models, which we show in an extensive evaluation.
Paper Structure (15 sections, 6 equations, 8 figures, 5 tables, 1 algorithm)

This paper contains 15 sections, 6 equations, 8 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Previous Work
  3. Meta-sampling
  4. Random
  5. Meta
  6. Meta-Learned sampling
  7. Models
  8. Implementation
  9. Evaluation
  10. Methodology
  11. Qualitative results
  12. Quantitative results
  13. Ablations
  14. Discussion
  15. Conclusion

Figures (8)

  • Figure 1: Projection.
  • Figure 2: Result for all methods (horizontal) and models (vertical) at equal sample count for the test-set BRDFs blue-rubber (top) and silver-paint (bottom) at $\TextOrMath{$n$\xspace}{n}=8$ and $\TextOrMath{$n$\xspace}{n}=32$, respectively.
  • Figure 3: Performance (vertical, log scale) of different learning methods (colors) for different models according to different metrics (Log. MAE in BRDF space and image-based DSSIM in every pair, lower is better) depending on the sample count (horizontal, log scale). The red dot indicates the theoretical optimum, when giving the model five orders of magnitude more samples, i.e., compute and acquisition time.
  • Figure 4: Test set loss (vertical) per BRDF (horizontal), sorted based on the results of Ours in decreasing order.
  • Figure 5: Results on data from nielsen2015optimal at $\TextOrMath{$n$\xspace}{n}=2$ samples.
  • ...and 3 more figures