BIESEL@RED.RUTGERS.EDU.UUCP (03/27/87)
I would appreciate pointers and references to current work in object recognition. My group is beginning work in the automation of visual database design for real-time image generators. These databases consist of polygonal approximations of real-world objects (everything from houses to bushes). Currently, individual objects are constructed by hand, using models, maps, photographs, graph paper, geometry, and lots of time and patience. We would like to develop a modeling station which can extract the basic geometry of objects from sets of photographs, and which can produce good approximations to polygonal models of the regular structures, such as buildings and other cultural features, which it recognizes in the source photographs. We expect that such a system will require some operator assistance for resolving ambiguities, at least initially, but even such a system would be of great help in the modeling task. Although we have some papers of current work, please assume that we are completely ignorant about who is doing what, and what the state of the art is, and forward all references to me. I realize that there are probably several netlists which are relevant, but I've not kept in touch with these. Pointers to the more active and relevant of these are also appreciated. I will summarize the responses if they are sufficiently general for this audience, and if the volume of replies warrants it. Many thanks in advance. Heiner BIESEL@RUTGERS [The best collections of papers are the DARPA Image Understanding Workshops. The February '87 proceedings have been made available to the general public. Much of this work is oriented toward aerial cartography (as well as target recognition). Other good papers have appeared in recent vision conferences such as PRIP/CVPR/ICCV and in journals such as IEEE PAMI and CVGIP. Some of the most pertinent work is being carried out at SRI by Pascal Fua and Andy Hanson. They have developed ways of extracting rectilinear objects (i.e., buildings of complex shape) and are extending their techniques to identify roads and vegetation. One of the inputs to their system is a segmentation map derived from my own work in computer vision. -- KIL]