Keypoint Matching with 3D3P Histogram Voting for Outlier Removal

Abstract

For removing outliers in keypoint matching, we develop a simple method by extending the existing method of histogram voting. With the spatial statistic constraint of the small camera motion assumption, correctly matched keypoints should yield displacement of a well-defined distribution in both image coordinates and keypoint orientation (3D). With the 3D displacement statistics, a histogram is built, in which three peaks (3P) instead of one are picked for selecting inlier candidates. The method shows better performance than traditional RANSAC or previous histogram voting methods in cases of both significant illumination changes and view-angle changes.