Course Introduction 

Lecture Notes 


Intro to Course (Jan 11)
[slides]
[6 per page] 


Review of prob and statistics (Jan 13)
[notes] 

Homework Assignments 


Homework 1 due Mon, Jan 24 

Reference Material 


Paper Search/Reading/Writing/Vision Resources 


Prob/Stats review chapters [on Angel course site] 

Gaussian Mixture Models and Expectation Maximization 

Lecture Notes 


The Gaussian Distribution (Jan 18,20,25)
[scanned notes] 


Intro to Gaussian Mixture Models (Jan28)
[scanned notes] 


GMM / EM derivation (Jan28,Feb1)
[slides]
[6 per page] 

Homework Assignments 


EM programming assignment due Mon, Feb 14 

Reference Material 


Intro to Gaussian Distribution , Bishop, PRML book 


Old and New Matrix Algebra Useful for Statistics, Tom Minka. 


Estimating Gaussian Mixture Densities with EM  A Tutorial, Carlo Tomasi. 


Mixture Models and EM , Bishop, PRML book 

EM Project Ideas 


 One straightforward idea for testing your EM implementation is to go back to the programming example from assignment 1 where you were plotting contours of the bivariate histograms for redgreen, greenblue and redblue color channels for images with "interesting" colors. Instead of a histogram representation, how about using EM to fit a mixture of Gaussian distribution to each 2D set of color pixel data (or even to the full 3D RGB color distribution). Think about how you might use the histograms you estimated earlier to check the sanity of the mixture of Gaussian distributions you estimate this time (for example, are significant modes found in the same places in both representations).



 From Bishop book: use kmeans or EM to label/compress colors in an image.
Try diff values of K. maybe compare kmeans and EM results. Bishop's idea text is
available here



 From Forsyth and Ponce: a number of ideas, including color/texture
segmentation (e.g. blobworld), fitting line segments (maybe compare EM with RANSAC?),
motion segmentation, background subtraction (Power and Shoonees paper). See the
[Forsyth and Ponce ideas].



 Two other project ideas from the last time I taught this course. One
involves motion segmentation (I provide a dataset), one involves experimenting with
the Jones and Rehg skin color classifier.
[here are the ideas].



 Of course, I encourage you to come up with your own ideas
involving data that is relevant to you.


Procrustes Analysis 

Lecture Notes 


Intro to Procrustes Analysis (Feb 1723)
[scanned lecture notes] 


Shape Models and PCA (Mar 1)
[slides]
[6 per page] 

Homework Assignments 


Procrustes analysis assignment due Thurs, Mar 17 


Bookstein's Schizophrenia Dataset 

Reference Material 


Solving for 2D similarity using complex numbers 


Cootes and Taylor Active Shape Model paper 


Fun Reading:
The Shape of Madness, Mackenzie, Discover Magazine 

Graphical Models 

Lecture Notes 


Intro to Graphical Models (Mar 22,24)
[slides]
[6 per page] 


Hidden Markov Models (Mar 29)
[slides]
[6 per page] 


Kalman Filter derivation (Mar 31)
[scanned notes]
These are lecture notes from my tracking course. Don't freak out when you see a reference to "homework" on the very first page! 

Reference Material 


PRML Graphical Models Chapter, Bishop, PRML book 


HMM Reading:
A Tutorial on Hidden Markov Models, Rabiner 

Markov Random Fields (Graphical Models II) 

Lecture Notes 


Brief Intro to MRF
[scanned notes] 

movie 
Dan Huttenlocher
Speeding Up Belief Propagation. There are several options of how to play it. I have verified that both "Stream on RTSPenabled devices" and the "QuickTime: 1084610846QuickTime.mov" link on that page work on my Windows machine (right clicking on the quicktime link lets you download and save the whole movie to watch locally). 

Final Programming Project Ideas due Tues, May 3 

Sampling and Markov Chain Monte Carlo 

Lecture Notes 


Intro to Sampling and MCMC (Apr 12,14)
[slides]
[6 per page] 

Reference Material 


Intro to Monte Carlo Methods, D.J.C.MacKay 


An Intro to MCMC for Machine Learning, Andrieu, DeFrietas, Doucet and Jordan 


MCMC for Computer Vision, ICCV 2005 Tutorial, Zhu, Dellaert and Tu 
