Analysis and Interpretation of Multiple Motions Through Surface Saliency

Abstract

The problem of recovering the 3-D camera and scene structure has been intensively studied and is considered well understood. Starting with two images, a process of establishing point correspondences is usually followed by the estimation of epipolar geometry while also rejecting outlier matches, and finally by 3-D structure estimation. However, most existing methods tend to fail in the combined presence of noise and multiple motions, since no single constraint applies to the entire set of matches. Hence, image registration becomes a more challenging problem, as the matching and registration phases become interdependent. We propose a novel approach that decouples the above operations, allowing for separate handling of matching, outlier rejection, grouping and 3-D interpretation. Our method first determines an accurate representation in terms of dense velocities, segmented motion regions and boundaries, by enforcing only the smoothness of image motion, followed by the extraction of 3-D camera and scene geometry.

This research has been funded in part by the Integrated Media Systems Center, an NSF Engineering Research Center, Cooperative Agreement No. EEC-9529152, and by NSF Grant 9811883.