Title: Model Fitting and Optimisation in Ravl

Abstract

In the past few years there has grown a concensus concerning a good and
robust approach to parametric model fitting in 2D and 3D vision, of which
canonical examples are fitting a 2D homography to two images of a plane,
and computing the fundamental matrix between two views of a stationary
3D scene. Both problems are usually attacked by first computing corner
features, and selecting candidate corner matches between the images.
The major issue then to be addressed is model fitting in the presence
of a large fraction of outliers in the match data, sometimes 50% or maybe
even more. There are two stages involved:

 1) Use Random Sampling Concensus (RANSAC) to separate inlier data from
    outlier data.
 2) Apply robust least-squares to fit the model to the inlier data, on the
    assumption that the inlier data set computed by RANSAC will still be
    partially corrupted by bad data.

This combination yields a fast and robust approach to a wide range of
robust fitting problems, such as:

* Motion estimation
* Curve fitting
* Camera calibration
* Object recognition

I have implemented classes for Ravl that encapsulate these methods, and
example classes and programs which use them. I will demonstrate a mosaicing
program which stitches images and builds mosaics in real time.
The talk will cover a description of the algorithms as well as an
overview of the Ravl optimisation classes.