Radiometric methods for the extraction of shape from images, such as photometric stereo, make simplifying assumptions about the light transport effects underlying those images. Among the most common assumptions are absence of interreflections and Lambertian reflectance. These make radiometric shape reconstruction techniques unsuitable for many classes of common objects, including objects with glossy surfaces or concave shapes. Our goal is to construct an inverse rendering framework that can be used to reconstruct shape and reflectance properties without these assumptions. Towards this goal, we develop a versatile, shape-differentiable, Monte Carlo renderer, which can efficiently estimate the differentials of image intensity values with respct to BSDF and local shape parameters. We combine this differentiable renderer with stochastic optimization and surface reconstruction algorithms, to develop a pipeline that estimates a 3D mesh that best explains captured image measurements. We evaluate this pipeline in experiments using both simulated and captured image datasets, and show that it can accurately reconstruct complex reflectance and shape even in the presence of strong global illumination. Finally, we discuss future extensions towards enabling the application of our inverse rendering framework to measurements from a large variety of 3D sensing systems.

Thesis Committee:
Ioannis Gkioulekas (Chair)
Srinivasa Narasimhan

Srinvasan Seshan, Head, Computer Science Department
Tom M. Mitchell, Interim Dean, School of Computer Science

56 pages

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by [email protected]