Dominant plane detection for uncalibrated binocular vision

Current methods for planar segmentation of binocular images require either calibrating the cameras or having correct stereo correspondence a priori. Here are introduced robust and extendable frameworks, that bypass these requirements. Feature vectors represent image structure, and ideas from scale space theory are applied. Two basic frameworks are introduced, one requiring (possibly partially incorrect) stereo correspondence a priori. The other method requires no a priori correspondence at all, instead finds the dominant plane while simultaneously matching features. The methods result in sparse segmentations, and a denser segmentation can be found by warping using the estimated dominant projectivity. Experiments are performed on real images, and different choices of feature vectors, methods and parameters are investigated. Proc. International Symposium on Intelligent Robotic Systems (ISIRS’98), pp. xx-xx, January 10-12, 1998, Bangalore, India