David Nistér and Andrew Davison
This tutorial covers both theoretical and practical aspects of real-time algorithms to recover the motion of moving single or multiple cameras and the scene geometry they observe. The real-time constraint forces special approaches, and we will look in detail at two quite different but complementary techniques derived from the fields of structure from motion and simultaneous localization and mapping.
The tutorial will include live demonstrations of the systems discussed. Specific topics include: Feature detection and tracking, Random sample consensus (RANSAC), RANSAC with fast scoring. Bundle adjustment and iterative refinement, triangulation using different error metrics. Minimal structure from motion solutions, the 3-point and 6-point methods for pose, the 5-point method for relative orientation, 6-point methods, 7-point method, 8-point method. Sampson approximation and scoring in two and three views. Stitching coordinate frames. Dense multi-view stereo, Fusing depth maps. Simultaneous Localization and Mapping (SLAM) with a single camera.
Andrew’s Slides can be found here
Video with audio of David’s part of the tutorial in divx format, thanks to Jason Meltzer. The first few minutes are missing, and the picture quality is not the best, but you may find the audio useful as you go through the slides.
Video and audio of Andrew’s part, also in divx format, thanks to Jason Meltzer.
David Nistér and Henrik Stewénius
The minimal problems for pose and relative pose have
provided many challenging problems for the computer vision community. The
solution of many of these problems is much easier when using the right tools
from Algebraic Geometry. We will give a very brief introduction to algebraic
geometry and direct our course towards the construction of Gröbner bases and demonstrate
how they can be used for solving geometric problems in computer vision. We will
do our best to make the course mathematically self-contained and will present
the theory in an intuitive manner in order to quickly get to where it is
interesting from a computer vision standpoint.
As examples to illustrate the method we will demonstrate a few trivial
examples, an easy way to solve the 5 point calibrated relative pose problem as
this problem is well known and the previous solvers are rather complicated,
finally we will demonstrate our solver for the relative pose for the
semi-calibrated case of unknown focal length.
ˇ Henriks
thesis contains a lot of material related to this.
ˇ A matlab solver for the
minimal calibrated relative pose problems. (five points, two cameras)