Abstract
Robust fitting is a basic technique and has been widely applied in photogrammetry and remote sensing, such as geometric correction. As known, typical robust estimators (include M-estimators, S-estimators, MM-estimators, etc.) often fail when outlier rate is higher than 50%, even if the outliers are uniformly distributed. In this article, we propose simple yet effective estimators, called adaptive M-estimators (AM-estimators). They are still robust under 80% of outliers. The proposed AM-estimators are very important supplements of M-estimators. Different from M-estimators, we use a varying parameter (shape-control parameter) instead of the original constant parameter in the weight function. The shape-control parameter decreases along with iterations in iteratively reweighted least squares, namely, AM-estimators are optimized in a coarse-to-fine manner. We adapt the proposed estimators into classical remote sensing and photogrammetry tasks, including mismatch removal, camera orientation (or called perspective-n-point), and point set registration to demonstrate their powers. Extensive synthetic and real experiments show that AM-estimators are superior to M-estimators, S-estimators, MM-estimators, and RANSAC-type methods. The source code of AM-estimators will be publicly available at https://ljy-rs.github.io/web .
Highlights
IntroductionE RRORS play a very important role in remote sensing and photogrammetry
E RRORS play a very important role in remote sensing and photogrammetry. They can be classified into three groups, i.e., 1) random errors, 2) systemic errors, and 3) gross errors [1], [2]
We briefly review classic or recently developed robust fitting methods, including maximum consensus methods, robust estimators, and other related methods
Summary
E RRORS play a very important role in remote sensing and photogrammetry. They can be classified into three groups, i.e., 1) random errors, 2) systemic errors, and 3) gross errors (outliers) [1], [2]. Systemic errors are usually caused by data acquisition devices such as laser scanners and imaging sensors, which can be compensated by a calibration model in the preprocessing stage. Random errors are assumed to follow a normal distribution, and least-square (LS) estimators are the best estimate for random errors [3]. Gross errors are undetected mistakes that cause observations to be far from the true values. Manuscript received March 26, 2021; revised April 19, 2021; accepted May 1, 2021.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
