PDF, Englisch Download (29MB) | Nutzungsbedingungen

Abstract

This thesis is concerned with the acquisition, modeling, and augmentation of material reflectance to simulate high-fidelity synthetic data for computer vision tasks. The topic is covered in three chapters: I commence with exploring the upper limits of reflectance acquisition. I analyze state-of-the-art BTF reflectance field renderings and show that they can be applied to optical flow performance analysis with closely matching performance to real-world images. Next, I present two methods for fitting efficient BRDF reflectance models to measured BTF data. Both methods combined retain all relevant reflectance information as well as the surface normal details on a pixel level. I further show that the resulting synthesized images are suited for optical flow performance analysis, with a virtually identical performance for all material types. Finally, I present a novel method for augmenting real-world datasets with physically plausible precipitation effects, including ground surface wetting, water droplets on the windshield, and water spray and mists. This is achieved by projecting the realworld image data onto a reconstructed virtual scene, manipulating the scene and the surface reflectance, and performing unbiased light transport simulation of the precipitation effects.