Visual Modeling using Motion and Light (Visueel modelleren met behulp van beweging en licht)
Author:
Keywords:
PSI_VISICS
Abstract:
Extracting or reconstructing shape information from images or videos of the world around us through computer algorithms is one of the core problems in computer vision. In this dissertation, we explore two approaches to the shape reconstruction problem based on visual cues involving motionand light, respectively.The first approach operates on multiple overlapping images of a static scene taken from different viewpoints. It is shown that it is possible to gradually retrieve all the necessary information to construct visual models from uncalibrated images by automatically extracting features, matching them between consecutive views and robustly estimating multi-view relations. Unfortunately, an important, but often ignored, problem of this approach is that it breaks down in the case of a (dominantly) planar scene. The degeneracies involved are described in detail, and a complete approach to overcome them with the aid of model selection is presented. The maturation of these techniques was also the inspiration for the development of a vision-based mobile mapping system, where a ground-based vehicle with pre-calibrated cameras mounted on top records images as it drives through the streets. The recorded images are then used to determine the relative positions and orientations of the vehicle between consecutive poses. The resulting georeferenced images can be used to measure structures of interest, i.e., digital surveying.In the second approach, the shading on a surface is used to infer information about local surface properties (normal and albedo) and hence overall surface shape. By observing a Lambertian surface under different known illumination conditions, we can tease apart the components that contribute to the shading. A method is presented that can deal with complicating effects such as shadows and non-Lambertian reflections. The redundancy in a dense set of images is exploited by a generative model that properly identifies and discards measurements not consistent with the underlying Lambertian assumption. Building on this, an easy-to-use, portable system for the digitization and visualization of small cultural heritage artifacts is also presented.