A robust corresponding points for uncalibrated stereo images

There are many problems when we find matching points of stereo images using only the epipolar geometry. Therefore, we use a correlation technique to find an initial set of matched points, and then examine motion vectors to remove mismatched points among the candidates. Then the epipolar geometry can be accurately estimated using a well adapted criterion with the fundamental matrix. We also introduce a bucketing technique to optimize fundamental matrix, and apply F-guided matching method to find best matches for the images. The proposed algorithm has been widely tested with various scenes which have many repetitive patterns. The experimental results show that performance of our algorithm is better than that of RANSAC method.