4.7 Article

MSSF: A Novel Mutual Structure Shift Feature for Removing Incorrect Keypoint Correspondences between Images

Journal

REMOTE SENSING
Volume 15, Issue 4, Pages -

Publisher

MDPI
DOI: 10.3390/rs15040926

Keywords

structure shift feature; mismatch removal; image matching; 3D reconstruction; pose estimation

Ask authors/readers for more resources

In this paper, a novel robust feature called Mutual Structure Shift Feature (MSSF) is proposed to remove incorrect keypoint correspondences. The feature measures the bidirectional relative ranking difference for the neighbors of a reference correspondence, and combines spatially nearest neighbors with geometrically good neighbors to reduce the negative effect of incorrect correspondences. The proposed method is evaluated through extensive experiments on raw matching quality and downstream tasks.
Removing incorrect keypoint correspondences between two images is a fundamental yet challenging task in computer vision. A popular pipeline first computes a feature vector for each correspondence and then trains a binary classifier using these features. In this paper, we propose a novel robust feature to better fulfill the above task. The basic observation is that the relative order of neighboring points around a correct match should be consistent from one view to another, while it may change a lot for an incorrect match. To this end, the feature is designed to measure the bidirectional relative ranking difference for the neighbors of a reference correspondence. To reduce the negative effect of incorrect correspondences in the neighborhood when computing the feature, we propose to combine spatially nearest neighbors with geometrically good neighbors. We also design an iterative neighbor weighting strategy, which considers both goodness and correctness of a correspondence, to enhance correct correspondences and suppress incorrect correspondences. As the relative order of neighbors encodes structure information between them, we name the proposed feature the Mutual Structure Shift Feature (MSSF). Finally, we use the proposed features to train a random forest classifier in a supervised manner. Extensive experiments on both raw matching quality and downstream tasks are conducted to verify the performance of the proposed method.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.7
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available