RIVAGe Feedback during Visual Integration : towards a Generic Architecture

We consider the comparative study of visual process integration within either a biological system (i.e. the parieto-ventral and parieto-dorsal pathways of the cortical visual system in the primate) or an artificial system. Both systems deliver an estimation of [where], that is to say the motion and structure of the observed scene and of [what], i.e. the perceptual grouping and labeling of objects in the scene. The long term goal is to elaborate a common theory about precise questions in both neurosciences and algorithms and their architecture in artificial vision, including computer vision applications. Within this framework, the function and behavior of adaptive feedback mechanisms is a key point and on the leading edge of biological studies. The core of this idea is that visual processing is build around: 1) a first computational step allowing to pre-process the input information, provide initial estimates, generate hypotheses about which models to use, ? 2) a refinement step using iterative mechanisms of optimization of the visual perception. Such mechanisms occur, sometimes implicitly, in artificial vision processes. They are mainly related to such problems as the combination of visual attributes, computed from different sources and then fused for ``what'' and ``where'' perceptual tasks; the use of a-priori information, obtained from higher-level visual modules. These modules define models estimated from the data. These models are either given a-priori (e.g. rigidity, shape regularity, ..) or chosen thanks to object labeling obtained during the first computational step. The present work is a theoretical study in three steps: (a) a systematic analysis of existing results in neuro-science, (b) an interpretation of these results from the viewpoint of the variational approach widely used in computer vision (c) a specification of a simulation tool of parts of the visual cortex.