Project 2: Statistical and Neural Pattern Recognition

Project Statement

Gender Classification problem using Support Vector Machines

Design a classifier using SVMs for identification of gender from visual information. This project requires you to implement the following:
1. Choice of kernel/feature vector.
2. Formulation of the classification problem.
3. Optimization of the cost function.
4. Inference: Analysis of your solution, including critique on the choice of support vector, and why your algorithm perform, decision surface etc.

  • Use the dataset given in this page
  • This is a two class problem, but may not be linearly separable. So justify the choice of kernel or use the non-linearly separable version of SVM.
  • Use PCA to bring the data to a lower dimensional subspace if required.
  • For the gender classification problem, you are expected to formulate the costs functions yourself. YOU MAY USE CODE FOR SVM AVAILABE FOR DOWNLOAD ON THE INTERNET.

    • OPTIONAL PART: Face Recognition with SVM

    Redo Project 1 with the same features that you used (PCA/FLD/ICA/SFS/MEDA/Whatever) but using SVM tto learn the classification. Use any code available on the web for this part alone and cite them in your report . A Sufficiently rewarding credit will be given for doing this part (to be announced).

    Deadline and Submission Guidelines

  • You are expected to submit a report on OR before the final exam. As usual, requests for extended submissions will be entertained on an individual basis.
  • The report should cover design and inference details, including, design of the feature vectors, choice of kernel, optimization, analysis of the support vectors, and reasoning on why (or not) your classifier works.
  • You are not expected to submit the code. As usual, you are bound by the University's Honor Pledge to follow guidlines of this project.

    Resources

  • DATASET: [Click]. IMPORTANT: If you use this dataset for some publication, do refer to "A.M. Martinez and R. Benavente. The AR Face Database. CVC Technical Report #24, June 1998", and thank Narayanan Ramanathan for producing aligned cropped images. Images 1-150 are male subjects and the rest female.
  • Technical report at MERL on "Learning Gender with Support Faces"[Click]
  • Gender Classification of Human Faces [Click]
  • libSVM: C Code for SVMs (thanks to Ashwin Swaminathan) [Click]
  • Links to various Matlab SVM codes. [Click]

    • FAQ (suitably modified) from Shaohua's website

    Q: I have questions or doubts on this project, what should I do?
    A: 1) Send your email to Prof. Chellappa (rama AT cfar DOT umd DOT edu) and cc Aswin (aswch ZAT cfar ZOT umd ZOT edu).
    2) Visit us during office hours.