Precise real-time outlier removal from motion vector fields for 3D reconstruction

Finding the correct correspondences in an image sequence is a significant task for deriving 3D structure from motion. Most research has concentrated on extracting and matching salient feature points for correspondence. Block-matching has largely been disregarded due to its significant number of correspondence-outliers and its complexity. However, nowadays real-time hardware is available to obtain block-motion vectors. We present a fast method to filter out more than 99.7% of all outliers and show that the obtained correspondences can be used to derive the 3D scene depth of real image sequences.