COMVIS: A communication framework for computer vision

We describe a general approach to integrate the information produced by dif ferent visual modules with the goal of generating a quantitative 3D reconst ruction of the observed scene and to estimate the reconstruction errors. The integration is achieved in two steps. Firstly, several different visual modules analyze the scene in terms of a common data representation: planar patches are used by different visual modules to communicate and represent the 3D structure of the scene. We show how it is possible to use this simpl e data structure to share and integrate information from different visual m odalities, and how it can support the necessities of the great majority of different visual modules known in literature. Secondly, we devise a communi cation scheme able to merge and improve the description of the scene in ter ms of planar patches. The applications of state-of-the-art algorithms allow s to fuse information affected by an unknown grade of correlation and still guarantee conservative error estimates. Tests on real and synthetic scene show that our system produces a consisten t and marked improvement over the results of single visual modules, with er ror reduction up to a factor of ten and with typical reduction of a factor 2-4.