Abstract
In the talk I will review some of my works on Epipolar Geometry Estimation. I will start with the BEEM algorithm which puts RANSAC within the general framework of search/optimization methods. Suggestions for the various components of the algorithm will be given.
In the second part of the talk I will present our recent work on matching images of extremely difficult image pairs for which general purpose algorithms usually fail. Two types of such cases are urban scenes with repetitive structures (e.g., windows) and scenes with extremely wide baselines.
Instead of suggesting a new improved general purpose algorithm, we propose a deterministic pre-processing algorithm, whose output is a set of putative matches and prior probabilities. This data can be given as input to most state-of-the-art epipolar geometry estimation algorithms, improving their results considerably.
The algorithm was tested on three state-of-the-art epipolar geometry estimation algorithms (BEEM,BLOGS and USAC) yielding much better results than the original algorithms. This was shown in extensive testing performed on almost 900 image pairs from six publicly available datasets.
Finally we will describe our new algorithm on Epipolar Geometry Estimation Using Noisy Pose Priors. Smartphones are equipped with sensors which are able to estimate the internal and external calibration parameters of the images they take. It is therefore natural to study how these parameters can be used for Epipolar geometry estimation. The main challenge is that these parameters are quite noisy and therefore naïve methods cannot be used.
We introduce SOREPP, a novel estimation algorithm designed to exploit pose priors naturally. It sparsely samples the pose space around the measured pose and for a few promising candidates applies a robust optimization procedure. It uses all the putative correspondences simultaneously, even though many of them are outliers, yielding a very efficient algorithm whose runtime is independent of the inlier fractions. SOREPP was extensively tested on hundreds of real image pairs taken by a smartphone. Its ability to handle challenging scenarios with extremely low inlier fractions of less than 10% was demonstrated.