Comparison of Outlier Filtering Methods in Terms of Their Influence on Pose Estimation Quality

Abstract

Local feature matching is an important problem in many applications of computer vision. Matching the descriptors depending only on the distances is not enough since normal matches are always affected by outliers. Starting from this problem, we aim in this project to make a comparison between outlier filtering methods, Adaptive Locally-Affine Matching (AdaLam), and Lowe’s ratio test in terms of their influence on the pose-estimation quality and time consumption. AdaLam is a hierarchical method designed to effectively exploit modern parallel hardware for fast and accurate outlier filtering based on local affine motion verification with a sample-adaptive threshold. Lowe’s ratio test matches key points based on distance measurements by comparing the distance of the two nearest neighbors for identifying distinctive correspondences. We have also applied two methods to extract key points, SIFT and ORB, and studied their effect on the outlier filters. To perform experiments, two methods were used in the pose-estimation pipeline and the conclusion is based on the quality metrics of the computed transformation matrix between a pair of images. Images pairs for the dataset were constructed from the TUM RGB-D Dataset. We have demonstrated that SIFT is better than ORB in terms of the total number of key points generated. We have also shown that AdaLam is better than Lowe’s ratio in terms number of correct matches and speed.