Motion Averaging
A Framework for Efficient and Accurate Large-Scale Camera Estimation in 3D Vision
Tutorial at CVPR 2018
Time and Location: 8:30am to 12:30pm on 18 June 2018 in Room 155-DEF
Venu Madhav Govindu
Department of Electrical Engineering
Indian Institute of Science
Bengaluru 560012 INDIA
venug[at]iisc.ac.in
Abstract
In recent years there has been growing interest in large-scale 3D reconstruction using both RGB and depth cameras. The concomitant need for accuracy, efficiency and scalability in camera motion estimation is addressed by the framework of motion averaging. Given many relative motion estimates between pairs of cameras, motion averaging solves for the 3D motions of individual cameras. The efficacy of motion averaging has attracted research interest leading to significant theoretical and algorithmic maturity. Owing to its major advantages over conventional approaches, motion averaging is now utilised in many 3D reconstruction pipelines. This tutorial will provide a comprehensive introduction to motion averaging in 3D vision. An intuitive and systematic understanding of the underlying geometry of matrix Lie groups will be developed. A comparative classification and summarization of various motion averaging methods will be presented. In addition, this tutorial will provide an exposition of algorithms and best practices. Along with developing a clear understanding of the state-of-the-art, this tutorial will aim to enable researchers to utilise motion averaging principles in novel contexts of large-scale structure-from-motion as well as dense 3D modeling using depth cameras.
Brief Biosketch of the Presenter
Venu Madhav Govindu obtained his Ph.D from the University of Maryland, College Park, USA. He is on the faculty of the Department of Electrical Engineering, Indian Institute of Science, Bengaluru, India. His primary research interests lie at the intersection of geometry and statistical estimation. The current interests of his research group include robust approaches to motion averaging, large-scale structure-from-motion problems, high quality shape reconstruction using depth cameras and the application of motion averaging in geometric SLAM. Details of his research group are available here.
Outline of Presentation
Topic | Details | Time |
Introduction | Motivation, Problem statement, Application contexts |
1 hr 45 mins |
Theory and Formulation |
Geometric properties and structure of matrix Lie groups, Group structure and properties of SO(2),SO(3),SE(3), Averaging on Lie groups |
|
Intrinsic Methods | Averaging of relative motions, Theoretical issues, Distributed consensus methods (Weiszfeld algorithm etc.), Cost functions, Convexity issues, Hardness/difficulty of problems, Graph theory considerations |
|
Break (30 mins) | ||
Robustness | M-estimators, IRLS, l-1 methods, RANSAC variants, Loop consensus statistics |
1 hr |
Extrinsic Methods | Rank-based methods, Matrix completion, Theoretical issues, Algorithmic considerations, Comparison with intrinsic methods |
|
Rotation Averaging | Quaternion averaging, Intrinsic methods, Conjugate rotations, Applications in large-scale SfM |
|
Translation Averaging | Problem statement, Parallel rigidity theory and existence of solution, Optimization of cost functions, Comparison of methods, Applications in large-scale SfM |
|
Hierarchical SfM | Motion averaging and hierarchical representations of large-scale SfM datasets, Properties and advantages, Comparison with incremental bundle adjustment |
45 mins |
Euclidean Motion Averaging | Averaging of relative motions in SE(3), Theoretical considerations, differences wrt SO(3), application in large-scale SfM, multiview alignment of 3D scans |
|
Conclusion | Relation with motion estimation in SLAM, graph SLAM etc., Open problems, Future directions |
Relevant Publications of the Presenter
Relevant Publications
Theory Intrinsic Methods (including Rotation Averaging) Extrinsic Methods Robustness Translation Averaging Euclidean Motion Averaging Hierarchical MethodsOther Related Papers (including SLAM)