Detection of Curvilinear Regions J. Miteran (work done jointly with J. Matas) Abstract: Biederman (1985) proposed the Recognition-By-Components theory of object recognition arguing it explains the successful human identification of objects despite changes in the size, scale or orientation of the image. In the theory, objects are build from 3-D volumetric primitives called Geons. Dickinson, Bergevin, Pentland, Jain, and many other researchers tried to develop automatic recognition system based on the theory . However, their systems worked only in highly constrained domains (controlled lighting, simple objects, etc.) in which segmentation is often performed manually. Our objective is to automatically extract in a viewpoint and illumination invariant manner regions from real images whose pre-image is a generalised cylinder or a part of it (most geons are generalized cylinders). The 2D region - a projection of the generalized cylinder - will be defined by a curve s(t),width w(t) and intensity profile I(t,x) on the crossection. A cost function J (or, equivalently, a probability distribution P) will be defined on the space of triplets (s,w,I) and the detection of the regions posed as an optimization problem. In the first part of the talk, the motivation for the research will be discussed, focussing on the utility of curvilinear features multi-view wide-baseline matching and object recognition. Next, a 2D probabilistic model of a curvilinear region will be proposed its relation to 3D generalised cylinders clarified. Related literature will be reviewed and the novelty of our approach discussed. Finally, preliminary experiments using a simplified model of the curvilinear feature will be shown.