An Outlier Detection Method for Feature Point Matching Problem

We present an efficient outlier detection method for finding consistent matching between two sets of feature points. We first define a kind of distance between each pair of candidate assignments which measures the compatibility between them. Using this distance measurement, correct assignments are generally compatible with each other and thus tend to form a cluster with high density. Our aim is to detect this correct assignment cluster by adapting an outlier detection method. We first present a new inlier scoring method, called Degree-Distance Inlier Scoring (DDIS), in which we integrate both degree and distance simultaneously based on kNN graph. Then we detect correct assignments and achieve point matching using DDIS and greedy algorithm. We call it as Outlier Detection Point Matching (ODPM). At last, we propose a more robust point matching algorithm by rendering ODPM in an iterative way. Experimental results on both synthetic and real-world data show the effectiveness the proposed method.