djb@LAFITE.BELLCORE.COM.UUCP (04/11/87)
In a recent posting Dave Chassin writes: > I know there are some people who have already done some work in these areas, > but it has always amazed me how little is in fact published. I have NO, get > that, NO references relating to 3D reconstructions other than the following, There has recently been a lot of published work on stereo reconstruction from image pairs. The work appears for the most part in two bi-annual confer- ences: International Conference on Pattern Recognition, and the International Joint Conference on Artificial Intelligence. On alternate years there are US ver- sions entitled AAAI (American Association for Artificial Intelligence) and CVPR (Computer Vision and Image Process- ing). The situation was different in the mid-1970's when the only stereo work available existed in the form of a few hard-to-locate PhD theses and some NASA-sponsored work on the Mars rover. Part of the problem was the lack of availability of video frame grabber hardware and the high cost of quality low noise cameras. When you add the amount of computing needed to perform even the lowest level of image analysis (e.g. edge extraction) it is no wonder that little attention was given to multiple-image problems. > That is that you will need to create some sort of algorithm for generating a > 3D model from 2D information received by the cameras. Essentially the idea > is to analyse a pair of images, extract the boundary data, assemble a 2D > 'image' for each view, project the two images together into a 3D 'image', My thesis ("On Computer Stereo Vision with Wire Frame Models," University of Illinois, Urbana, January, 1978) addressed two different problems in stereo robot vision. The first problem was to automatically recon- struct a 3-D wire frame of an object from multiple views. The second problem was to align a pre-existing wire frame model (constructed by hand) to the partial 3-D wire frame derived from the images. I described a three-view stereo technique which corrected some of the existing problems associ- ated with binocular stereo. The three views are arranged in a triangle. This arrangement reduces matching ambi- guity and increases triangulation accuracy for edge com- ponents parallel to the disparity direction. I sug- gested that more than three views could be used to further enhance the signal-to-noise ratio. Others later applied a similar technique to successfully guide an autonomous robot vehicle. The second part of the system aligned a pre- constructed wire frame model of the object to the image- derived wire frame. It was entirely automatic and could even align partially occluded objects. This allowed a robot to "see" an object and grasp it. The document is available as a Technical Report from: Coordinated Science Laboratory, University of Illinois, Urbana, IL 17604 . The thesis is available from: University Microfilms, Ann Arbor, Michigan. A short paper appeared in the Fifth International Joint Conference on Artificial Intelligence, MIT, Cambridge, MA, 1977, co-authored by D. J. Burr and R. T. Chien, entitled "A System for Stereo Computer Vision with Geometric Models". D. J. Burr, 2A-397 Bell Communications Research 435 South Street Morristown, NJ 07960 email: djb@bellcore.com