Material appearance is a complex composite of its geometry, underlying physical model, and the illumination condition in which it is observed. Traditional approaches require dedicated and rather complicated setups to recovery both geometry and underlying physical model while controlling the lighting condition. Although results from traditional approaches, in general, are accurate, today’s digital content creation desires simple setups and rapid acquisition procedures, while material models ac- quired still produce plausible and photo-realistic images.
As alternatives to complex traditional approaches, novel acquisition methods and algorithms employ simplified acquisition procedures with more relaxed constraints than transitional approaches and less required data. Our work falls into such a category.
This thesis explores novel methods of rapid and straightforward material shape and reflectance acquisition, with the smallest possible number of images - one. We successfully demonstrate a shape-agnostic method that extracts isotropic and homogeneous material information, i.e., reflectance parameters and mesostructure variation, from a single image with known natural illumination using image statistics. We further extend the problem’s scope to non-homogeneous materials and present a novel method that combines deep learning and polarisation from reflection to estimate an exemplar’s reflectance and geometry with a flashlight.
Evaluation is an equally important part as the methods in this thesis. We evaluate the robustness and versatility of our approaches both quantitatively and qualitatively. We capture real-world material exemplars and qualitatively compare the rendered results with photographs. On the other hand, we perform synthetic tests and quantitatively analyze results against the ground truth. Finally, we perform comparisons against existing methods and demonstrate superior results.