Utilize feature distinctiveness to recover feature correspondences

Feature corresponding problem is important for many computer vision tasks, but it is very difficult when the feature sets are corrupted by noise features. This paper formulates this problem as an optimization problem, and then proposes to measure distinctiveness of one feature match based on appearance similarity between two features. Then candidate feature matches are initialized based on their distinctiveness values. By weighting each candidate feature match by its distinctiveness value, the feature corresponding map can be robustly estimated by weighted Support Vector Regression Machine. Then the outlier feature matches are rejected by checking their geometric consistence with the estimated corresponding map. The proposed algorithm iterate above steps until the true feature correspondences are recovered. Experimental results demonstrate the effectiveness of this method.