code-for-Lucas-Kanade-20-Years-On

Lucas-Kanade 20 Years On: A Unifying Framework

Part 1: The Quantity Approximated, the Warp Update Rule and the Gradient Descent Approximation

Simon Baker and Iain Matthews

The Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 USA

Contact: [email protected], [email protected]

URL: http://www.ri.cmu.edu/projects/project_515.html

Bibtex reference:

@article{Baker_2004_4293, author = "Simon Baker and Iain Matthews", title = "Lucas-Kanade 20 Years On: A Unifying Framework Part 1: The Quantity Approximated, the Warp Update Rule, and the Gradient Descent Approximation", journal = "International Journal of Computer Vision", year = "2004" }

                         MATLAB SOURCE CODE

The source code is split into two directories, both need to be added to your Matlab search path:

lk20-p1 The algorithms described in the paper and test scripts lk20-common Common code for all algorithms (and for later papers)

lk20-p1/

The Gauss-Newton affine warp algorithms in Section 3 of the paper are:

affine_fa.m Forwards-Additive affine_fc.m Forwards-Compositional affine_ia.m Inverse-Additive affine_ic.m Inverse-Compositional

The Gauss-Newton projecive warp algorithms in Section 3 of the paper are:

homo_fa.m Forwards-Additive homo_fc.m Forwards-Compositional homo_ic.m Inverse-Compositional

Note: the inverse-additive algorithm cannot be applied with a projective warp.

The additional algorithms from Section 4 of the paper are:

affine_ic_nt.m Inverse-Compositional Newton (4.2) affine_ic_sd.m Inverse-Compositional Steepest Descent (4.3) affine_ic_d.m Inverse-Compositional with diagonal Hessian (4.4) affine_ic.nt_d.m Inverse-Compositional Newton with diagonal Hessian (4.4) affine_ic_lm.m Inverse-Compositional Levenburg-Marquardt (4.5)

To run the perturbation experiments described in the paper examine:

run_affine.m Run an affine warp experiment run_homo.m Run a projective warp experiment

These files should be edited for the experimental parameters you would like to test. By default they run everything for a few tests. To replicate the experiments in the paper run each algorithm for 100 convergence tests and 1000 frequency of convergence tests. This will take some time!

lk20p1/data/

The data directory contains precomputed random perturbation data (for experimental consistency) and the images and template definitions.

lk20p1/figures/

The figures directory contains matlab scripts for plotting your results and creating graphs similar to those shown in the paper.