Deep image-based Adaptive BRDF Measure

TL;DR

This work tackles the slow, dense measurement of BRDFs by introducing an image-driven, adaptive sampling strategy. A lightweight CNN estimates Ward BRDF parameters from a material image and defines an adaptive sampling pattern via BRDF importance sampling, with an image-based loss to determine the required sample count. The approach is validated using Ward model syntheses and MERL data, demonstrating accuracy close to ground truth with significantly reduced measurements and outperforming a recent meta-learning sampling method. The method enables faster, high-fidelity BRDF capture suitable for rendering and sensor simulation, with potential extensions to other BRDF models and sampling techniques.

Abstract

Efficient and accurate measurement of the bi-directional reflectance distribution function (BRDF) plays a key role in high quality image rendering and physically accurate sensor simulation. However, obtaining the reflectance properties of a material is both time-consuming and challenging. This paper presents a novel method for minimizing the number of samples required for high quality BRDF capture using a gonio-reflectometer setup. Taking an image of the physical material sample as input a lightweight neural network first estimates the parameters of an analytic BRDF model, and the distribution of the sample locations. In a second step we use an image based loss to find the number of samples required to meet the accuracy required. This approach significantly accelerates the measurement process while maintaining a high level of accuracy and fidelity in the BRDF representation.

Figures (10)

• Figure 1: Rendered balls with ground truth vs our adaptive measurements
• Figure 2: Method Flowchart of Deep Image-based Adaptive Reflectance Measure. For a fixed material, we use its image as input to an encoder network, which then estimates the BRDF parameters of it. An adaptive sampler use these parameters to determine the outgoing direction locations. Finally, we progressively increase the number of these locations to achieve the minimum number of samples required while maintaining high fidelity.
• Figure 3: Encode Network Architecture: The blue boxes denote convolutional layers that are integrated with batch normalization and ReLU activation functions. The dimensions of these layers are the numbers inside them. A green box represents a fully connected layer outputting the BRDF parameters.
• Figure 4: A visualization of the process g in Eq. \ref{['brdf_eqa']} to calculate the adaptive sampler's position. We start from get sample points $\left(u_1, u_2\right)$ on a uniform grid in the unit square $[0,1]^2$. The importance sampling process takes a sample point $\left(u_1, u_2\right)$, and maps it to position on a 2D BRDF slice ,and reverse vise works by its inverse function.
• Figure 5: Predicted Values from BRDF Nerual Network VS Ground Truth.