This Ph.D. dissertation addresses the problem of estimating depth ordering information from single images, a key issue in image understanding that in recent years has focused the interest of the community. Motivation behind this tendency is provided by several important applications that could beneficiate of advances in the field such as automatic object removal, image indexing, 3D scene reconstruction and synthesis.
In contrast to state-of-the-art works, this Ph.D. dissertation investigates a general low-level approach to the problem of monocular depth estimation, in which the depth ordering is directly inferred from a set of monocular depth cues without relying on any previously learned contextual information nor on any assumption on the image structure. New methods for monocular depth cue detection are proposed, the problem of depth cue integration is analysed, and depth ordering information is exploited to improve classical color segmentation, and create new depth-oriented fitering applications. The investigation for depth cue integration leads to the development of two distinct frameworks based on different strategies: a diffusion based strategy and a region merging based strategy.
The former is based on the use of a nonlinear filter which iteratively extends initial depth values arisen from monocular depth cues to the entire image domain until stability is attained.
The result is a flexible framework that allows the integration of several monocular depth cues and that gives a correct interpretation of a plurality of vision phenomena.
The latter strategy is based on the construction of a hierarchical region-based representation of images, that incorporates depth ordering information provided by depth cues, as well as on a graph formalization, which encodes depth relationships between regions and allows to infer a global, consistent depth ordering.
© 2001-2024 Fundación Dialnet · Todos los derechos reservados