Image Matching via Ranking and Neighborhood Manifold Embedding

Image matching, which aims to identify reliable correspondences between related images, plays a fundamental and crucial role in many visual tasks. This article focuses on the mismatch removal problem for robust point feature-based matching. First, a robust and effective neighborhood manifold embedding strategy is proposed to preserve the underlying neighborhood topologies of potential true matches. Specifically, we introduce a motion-consistency constraint for neighborhood construction and preserve the manifold of an inlier through finding an optimal subspace embedding fit by its neighbors. Second, a neighborhood element ranking consensus for neighborhood refinement is developed to enhance feature matching performance in the presence of high outlier ratios. We term the combination of these two strategies as ranking and neighborhood manifold embedding (RNME) and further formulate the mismatch removal problem into a mathematical optimization model with a closed-form solution. Extensive experiments of general feature matching and image registration on various challenging datasets demonstrate the superiority of RNME over the current state-of-the-art approaches. Moreover, the method also exhibits favorable potential in 3-D reconstruction tasks.