# CALTECH EE 150 - Homework 2 (2 pages)

Previewing page 1 of 2 page document
View Full Document

## Homework 2

Previewing page 1 of actual document.

View Full Document
View Full Document

## Homework 2

44 views

Pages:
2
School:
California Institute of Technology
Course:
Ee 150 - Topics in Electrical Engineering
##### Topics in Electrical Engineering Documents
• 2 pages

• 22 pages

Unformatted text preview:

EE150 Homework 2 Due Thurs April 14 Before Class 1 Find a Project Several project descriptions have been posted Please look at these by Tues April 12 Remember the main portion of this class will be spent implementing and working on your project Please spend several hours looking over the papers and the project descriptions to identify projects that interest you 2 Eigenfaces vs Fisherfaces Read the Paper Belhumeur et al Eigenfacs vs Fisherfaces Recognition Using Class Specific Linear Projection located on the website It presents a different view on both PCA and Fisher from that which was presented in class Both of these techniques occur frequently within vision and pattern classification applications 1 General Fisher vs PCA questions Describe in your own words the subspaces which Fisher and PCA project to Fisher is a supervised technique It requires an extra piece of information associated with each data point which PCA does not necessarily require What is this extra information 2 Why does the image for a face lie in a 3D linear subspace if we consider it a Lambertian surface 3 Why do the authors achieve better performance when they ignore the first 3 principle components What do these components represent Would you expect the same improvement if you removed the first 3 Fisher components 3 Male and Female Faces In class we suggested the challenge of achieving the highest performance on Male vs Female faces the data set is online via the class website You have been given code which replicates the experiments which were presented in class You have also been given the data sets which were used Try to obtain the best possible classification results 1 1 What is over fitting When do we observe over fitting Assume we are fitting a high dimensional parametric model to our data Draw a plot of how the test and train performance would vary as a function of the number of training examples Explain your plot Describe or show how these curves would change if the amount of over fitting were increased or decreased 2 Try at least 3 different techniques which you feel are reasonable to obtain the best possible classification performance on Males vs Females Use portions of the code from class if you wish as a framework for your own code 3 These approaches Fisher and PCA are often termed global approaches to image recognition and classification They are not well suited for handling occlusion and or background clutter Test how well your system performs if you occlude part of the image Specifically create a plot indicating performance as a function of the amount of image occluded Only occlude images in the test set It is up to you how to generate the occlusion but it should vary from image to image within the test set Show us some of the occlusions you generate 2

View Full Document

Unlocking...