EXTRACTION OF 3-DIMENSIONAL STRUCTURE FROM IMAGES OF MULTIPLE VIEWPOINTS BY 2-STEP VOTING

This paper proposes a method to extract 3-D information from images of multiple viewpoints, avoiding the problem of feature point matching w hich is common in stereo techniques. A straight line connecting a char acteristic point in an image and the lens center (a backprojection lin e) goes through a real point on the object. Thus, the object space is divided into small voxels. For each backprojection line of all images, give a vote to each voxel which intersects with it and extract voxels with high score. Then they are expected to be common to many images a nd represent characteristic points of the object in 3-D space. However , in actual calculation, it is not so simple. For example, there are p roblems of: (1) voxel size selection; (2) interference of very near po ints; and (3) difference of the scores caused by occlusion. (Details a re mentioned later.) Problem (1) will be solved by introducing a conce pt that voxels are supported by images instead of backprojection lines . Problems (2) and (3) will be solved by two-step voting: the first vo te is done uniformly along the backprojection lines, and the second is done selectively based on the result of the first vote. Since the pro posed voting method is not affected by the voxel size, it is possible to adopt an algorithm to begin with large voxels and divide only neces sary voxels into smaller ones recursively until necessary resolution i s gained. Experiments with synthesized images and actual images show v alidation of the method.