SVBRDF Bootstrapping

SVBRDF bootstrapping is a new method for data-driven modeling of real-world, spatially-varying reflectance that provides a high resolution result in both the spatial and angular domains. It decomposes reflectance measurement into two phases. The first acquires representatives of high angular dimension but sampled sparsely over the surface, while the second acquires keys of low angular dimension but sampled densely over the surface. Because reflectance is spatially coherent, keys can be used as feature vectors to “unlock” or interpolate between representatives, using a linear combination of a small number of representatives (k=15) to match each key.

We develop a hand-held, high-speed BRDF capturing device for measurements in the first phase. A condenser-based optical setup collects rays emanating from a single point on the target sample over a dense hemisphere of outgoing directions. The device includes 6 LED light sources and captures the sample’s reponse to each light direction in sequence, yielding 10 BRDF point measurements per second. We then scan this device over the target to capture representatives at different, indiscriminate points. Sparse light measurements are amplified to a high-resolution 4D BRDF using a general microfacet model.

The second phase captures images of the entire sample from a fixed view. Lighting is varied by scanning a point source over the sample, yielding N=20-60 images from different light directions. Arbitrary background lighting in the enviroment may also be present and is measured and accounted for. The key measurement thus represents the material’s response to N key lighting conditions at a dense set of key points. We show that the resulting, N-dimensional response captures much of the distance information in the original BRDF space. At each key point, the same linear weighting computed to match the measured key response is applied to the full 4D representatives, yielding a high-resolution SVBRDF. A few minutes of capture on a simple device yields sharp and anisotropic specularity and rich spatial detail.