Voting-Based Approach to Nullspace Search for Correspondence Matching and Shape Recovery

Abstract

A simultaneous search, callednullspace search, for matching correspondences among images and recovering 3-D objects is proposed by using a voting-based method to circumvent erroneous recovery of 3-D objects arising from wrong matched correspondences among images. It is able to avoid occlusion problems and cope with remarkable changes in visibility in a long image sequence. An experiment is done with synthetic and real image sequences, consisted of 30 images of a sphere and 10 images of a toy house, under the condition that 3-D points are occluded at most 50% of the sequence and the camera moves with rotational as well as translational motions. The proposed method gives a basis for organizing multiple dynamic images where occlusion occurs frequently.