Vision and Learning CIS680 Project

Object Recognition Challenge:

Develop an object recognition algorithm with the following test cases in mind. You may assume the image to be recognized is segmented.

1. NIST digit database, 10 object class, 60,000 training/10,000 testing cases

http://yann.lecun.com/exdb/mnist

You need not to use the entire 60,000 training image, but you need to report error rate on the 10,000 test image.

2. CMU Face recognition database, 2 object class,

http://www.cis.upenn.edu/~jshi/Teaching/cis700-002/Project/FACE/

The algorithm should use 75% of the faces, and non-faces in the database as training example, and report ROC curve on the remaining 25% in the database.

You need to write the following three functions:

Training.m

Input:

X: training images

Y: class labels for the corresponding training images

Para: additional parameters for the learning alorithm

Output:

H: a classifier with its parameters that maps X-> Y

Testing.m

Input:

X: test images

H: a classifier with its parameters that maps X-> Y

Note: if you need to use the nearest neighbor classifier, you should store the necessary information in H.

Ouput:

Y: class label for each of the test images

Error_measure.m

Input:

Y_target: correct class labels

Y_computed: computed class labels

Options: the type of error need to be comptued

Output:

Test_error:

ROC:

Possible Algorithms:

You are free to develop any algorithm you wish for this task.

• Adaboosting.

You can choose a number of classifier to boost. Hand build feature (such as Viola face feature), decision tree.

Viola&Jones: http://citeseer.ist.psu.edu/viola01robust.html

• Support Vector Machine. Both the linear version and the nonlinear kernal version.

Osuna: http://citeseer.ist.psu.edu/osuna97training.html

• Hidden Markov, Markov Random Field.

Ioffe&Forsyth: http://www.cs.berkeley.edu/~ioffe/trees.pdf

Felzenszwalb&Huttenlocher: http://www.cs.cornell.edu/~dph/papers/pictorial-structures.pdf

• Feature Alignment, Statistical features.

Belongie, Malik, Puzicha: http://citeseer.nj.nec.com/belongie02shape.html

Scheinderman&Kanade: http://www.cs.cmu.edu/afs/cs.cmu.edu/user/hws/www/IJCV_final4.pdf

• Combining multiple simple classifiers with max. entropy formulation.

Mahamud, Hebert, Lafferty: http://www.cs.cmu.edu/~mahamud/eccv-2002.pdf

In design such algorithms, there are several issues you need to make a decision on.

• Image features. Hand designed or self-discovered with large data set. For this set of test, you may choose to use edge detection, filter output (as in the texture synthesis), Shape Context, Viola’s face feature.

• Geometrical alignment. To-do or not-to-do.

• How do model multi-class classification.

Deadlines:

• April 8, 2004. 2 pages write up on your algorithm design, implementation plan, and running time analysis of the learning algorithm, and of the classifier.
• April 22, 2004. 15 presentations on your algorithm and test results.
• May 5, 2004. Final write up, 8 pages with test results.