Evolutionary Optimization for Robust Epipolar-Geometry Estimation and Outlier Detection

Mozhdeh Shahbazi,Gunho Sohn,Jérôme Théau

doi:10.3390/a10030087

Mozhdeh Shahbazi, Gunho Sohn + Show 1 more

Open Access

https://doi.org/10.3390/a10030087

Copy DOI

Abstract

In this paper, a robust technique based on a genetic algorithm is proposed for estimating two-view epipolar-geometry of uncalibrated perspective stereo images from putative correspondences containing a high percentage of outliers. The advantages of this technique are three-fold: (i) replacing random search with evolutionary search applying new strategies of encoding and guided sampling; (ii) robust and fast estimation of the epipolar geometry via detecting a more-than-enough set of inliers without making any assumptions about the probability distribution of the residuals; (iii) determining the inlier-outlier threshold based on the uncertainty of the estimated model. The proposed method was evaluated both on synthetic data and real images. The results were compared with the most popular techniques from the state-of-the-art, including RANSAC (random sample consensus), MSAC, MLESAC, Cov-RANSAC, LO-RANSAC, StaRSAC, Multi-GS RANSAC and least median of squares (LMedS). Experimental results showed that the proposed approach performed better than other methods regarding the accuracy of inlier detection and epipolar-geometry estimation, as well as the computational efficiency for datasets majorly contaminated by outliers and noise.

Highlights

Sparse image matching is one of the most critical steps in many computer vision applications, including structure from motion (SfM) and robotic navigation
We focus on the problem of outlier detection based on the robust estimation of epipolar geometry
The experimental results obtained from the proposed technique are compared with those of the following state-of-the-art techniques: Random sample consensus (RANSAC), m-estimator sample consensus (MSAC), Maximum likelihood estimation sample consensus (MLESAC), LO-RANSAC (Lebeda et al, 2012), StaRSAC, Cov-RANSAC, Multi-GS-RANSAC and least median of squares (LMedS)

Summary

Introduction

Sparse image matching is one of the most critical steps in many computer vision applications, including structure from motion (SfM) and robotic navigation. Given the recent advancements in the fields of low-altitude, oblique and ultra-high resolution imagery, the rate of contamination has increased, and detecting the correct correspondences with high accuracy has become more challenging [6]. This is due to several factors, which include noisy measurements, the inefficiency of local descriptors, the lack of texture diversity and the existence of repeated and similar patterns that cause matching ambiguity [7,8,9]. The term inlier applies to true matches among the putative correspondences, and the term outlier refers to false matches.

Methods

Results

Conclusion