Abstract

Estimating material properties and modeling the appearance of an object under varying illumination conditions is a complex process. In this thesis, we address the problem by proposing a novel framework to re-illuminate scenes by recovering the reflectance properties. Uniquely, following a divide-and-conquer approach, we recast the problem into its two constituent sub-problems.

In the first sub-problem, we have developed a synthetic dataset of spheres with realistic materials. The dataset has a wide range of material properties, rendered from varying viewpoints and under fixed directional light. Images from the dataset are further processed and used as reflectance maps used during the training process of the network.

In the second sub-problem, reflectance maps are created for scenes by reorganizing the outgoing radiances recorded in the multi-view images. The network trained on the synthetic dataset, is used to infer the material properties of the reflectance maps, acquired for the test scenes. These predictions are reused to relight the scenes from novel viewpoints and different lighting conditions using path tracing.

A number of experiments are conducted and performances are reported using different metrics to justify our design decisions and the choice of our network. We also show that, using multi-view images, the camera properties and the geometry of a scene, our technique can successfully predict the reflectance properties using our trained network within seconds. In the end, we also present the visual results of re-illumination on several scenes under different lighting conditions.