DNF: Feature Point Matching Pairs Filter Based on Descriptor Net

Wide-baseline image registration under out-of-plane rotation and larger viewpoint change is still challenging. Most of the commonly used matching algorithms are not invariant to affine transformation. They heavily rely on the local features of image patches and ignore global information, the mismatch is inevitable and greatly affect the accuracy of image registration. To address this issue, we propose a feature point matching pair filter based on global spatial position correspondences of feature points, coined Descriptor Net Filter (DNF). We put forward two criteria to evaluate matching quality. One is the local matching quality computed by independent local feature, the other is the global matching quality relying on geometric network constraint. Combining the advantages of both local feature and large-scale geometric constraint, our method removes mismatches effectively. The experiments on both planar scenes and 3D objects from several standard datasets show that the DNF significantly enhances the matching precision and retains more correct matches as well.