Efficient Two-view Estimation using Richer Geometric Correspondences

Jonathan Astermark will present his licentiate thesis with the title

Efficient Two-view Estimation using Richer Geometric Correspondences

Abstract:

Two-view estimation is a fundamental problem in 3D computer vision, and an important sub-task of multi-view estimation pipelines such as Structure-from-Motion (SfM) and Simultaneous Localization and Mapping (SLAM). In recent years, the main focus in the field has been on keypoint-based methods, where interest points are first detected and matched across the two images, followed by robust estimation of the geometry based on these keypoints. In this robust estimation step, the traditional approach is to use keypoint coordinates as basis for the estimation, by minimizing a geometric residual such as the reprojection error. 

This thesis investigates estimation based on keypoints using richer geometric information, in addition to the keypoint coordinates. Thegoal is to increase efficiency in the estimation to achieve better runtime with maintained accuracy, which is an important factor for inclusion in multi-view systems for SfM and SLAM. 

The thesis is based on three papers; the first two concern sample efficient minimal solvers for relative pose and homography estimation, using keypoints augmented with additional geometric information. In the first paper, we use scale information to constrain relative depths when estimating relative pose. In paper II, we use both scale and orientation information to constrain estimation of a plane-induced Euclidean homography. By combining multiple similar and seemingly redundant constraints, we develop a novel minimal solver allowing us to get noisy but surprisingly good homography estimates from even a single correspondence. The third paper is focused on summarizing semi-dense keypoint matches, to harness recent improvements in dense, detector-free keypoint matching. We introduce a summarization scheme that reduces the redundancy of semi-dense keypoints, which significantly decreases runtime compared to traditional estimation, with negligible reduction in estimation accuracy.

The thesis is available at https://vision.maths.lth.se/jastermark/thesis/licentiate_thesis.pdf

Opponent: Professor Juho Kannala, Aalto University

Examiner: Associate professor Andreas Langer, Lund University

Supervisors: Anders Heyden and Viktor Larsson