Early Cognitive Vision: Feedback Mechanisms for the Disambiguation of Early Visual Representation
Nicolas Pugeault · 2008
Recent years have seen considerable progress in low level image processing as well as model based, vision applications. Joining the two fields proves to be a difficult problem due to the local ambiguity and noisiness of visual processes, and to requirements of robustness and accuracy of high level processes. In this thesis we present an Early Cognitive Vision framework that aims at providing a rich and reli- able scene representation from visual information. This framework preserves conflicting hypothesis in its early stages, and makes use of feedback mechanisms between different visual processes and layers of representation to achieve disambiguation. In a first part we present how a symbolic local image de- scriptors are extracted from the responses of early vision filters, and how perceptual grouping constraints can be applied on the resulting image representation. A second part discuss the use of stereopsis to reconstruct an equivalent 3D representation of the visual information. Interactions between perceptual grouping, stereopsis and 3D reconstruction processes are discussed. The third part of this thesis integrates visual information across time to further disambiguate the representation. This framework has been used successfully in several contexts, that are discussed in the conclusion of this thesis.