In this paper, we use local feature transformations estimated in the matching process as initial seeds for 2D homography estimation. The number of testing hypotheses is equal to the number of matches, naturally enabling a full search over the hypothesis space. Using this property, we develop an iterative algorithm that clusters the matches under the common 2D homography into one group, i.e., features on a common plane. Our clustering algorithm is less affected by the proportion of inliers and as few as two features on the common plane can be clustered together; thus, the algorithm robustly detects multiple dominant scene planes. The knowledge of the dominant planes is used for robust fundamental matrix computation in the presence of quasi-degenerate data.
展开▼
