An Evaluation of Feature Matchers Forfundamental Matrix Estimation.
Jia-Wang Bian,Yu-Huan Wu,Ji Zhao,Yun Liu,Le Zhang,Ming-Ming Cheng,Ian Reid
DOI: https://doi.org/10.5244/c.33.89
2019-01-01
Abstract:Matching two images while estimating their relative geometry is a key step inmany computer vision applications. For decades, a well-established pipeline,consisting of SIFT, RANSAC, and 8-point algorithm, has been used for this task.Recently, many new approaches were proposed and shown to outperform previousalternatives on standard benchmarks, including the learned features,correspondence pruning algorithms, and robust estimators. However, whether itis beneficial to incorporate them into the classic pipeline isless-investigated. To this end, we are interested in i) evaluating theperformance of these recent algorithms in the context of image matching andepipolar geometry estimation, and ii) leveraging them to design more practicalregistration systems. The experiments are conducted in four large-scaledatasets using strictly defined evaluation metrics, and the promising resultsprovide insight into which algorithms suit which scenarios. According to this,we propose three high-quality matching systems and a Coarse-to-Fine RANSACestimator. They show remarkable performances and have potentials to a largepart of computer vision tasks. To facilitate future research, the fullevaluation pipeline and the proposed methods are made publicly available.