| Summary: | 碩士 === 華梵大學 === 資訊管理學系碩士班 === 93 === Robust estimation was widely used in computer vision. The commonly-used applications in computer vision, such as: line fitting, conic fitting and fundamental matrix. This paper focuses on the analysis and comparison of some popular robust estimators, for example, M-estimator, LMedS (least median of squares), RANSAC (RANdom SAmple Consensus), and MINPRAN (MINimize the Probability of RANdomness). According to our experiments and analysis, we found that there are some problems in MINPRAN. Therefore, in this paper, a new method, Modification of MINPRAN, was proposed to solve some problems found from MINPRAN.
The applications, such as line fitting and fundamental matrix, were used in our experiments. In the experiment of line fitting, 50% outliers are set in sample data. The experimental results reveal that Modification of MINPRAN performs best; in experiments of estimating the fundamental matrix, 33% outliers and 60% outliers are set respectively. In the case of 33% outliers, LMedS achieves the best result, and Modification of MINPRAN performs well too; in the case of 60% outliers, LMedS definitely fails, but Modification of MINPRAN still may have 40 % correct results. Besides Modification of MINPRAN, a simplified vision of MINPRAN, called as Voting, was proposed.
Three methods to verify experimental results are used in this paper. 1. selected 8 corresponding point pairs are all inliers or not; 2. the mean value of residuals; 3. selected 8 corresponding point pairs on a plane or not (the concept of the homography is used). In fundamental matrix, because of degeneracies problems, the homography was used to calculate a value H in order to avoid the coplanar problems. According to our experimental results, Voting has the minimum mean of residuals and the largest H value.
In order to improve the flexibility of Voting, automatic determination of thresholds will be studied. In the future work, we will try to focus on the research of adaptive-scale robust estimators.
|
|---|
