ECE 281b: Final Project

1
ECE 281b Final Project

• Mei-Chen (Mei) Yeh
• 05/31/2006

2
Bush vs Schwarzenegger
3
Main goal

• Face categorization problem
• Object class recognition techniques have seen
  progress in recent years (ex bag-of-words
  models)
• Different persons are considered different object
  classes
• Integration of SIFT features
• Does SIFT help face recognition? To what degree?
• Target on few people
• Applications
• Face annotation in family albums
• Name-based photo search

4
Datasets

• The BioID Face Database (simple)
• 1521 images with 23 people
• Variety of illumination, background and face size

5
Statistics of the BioID Database
6
Datasets

• http//www.cs.berkeley.edu/millert/faces/faceDict
  /NIPSdict/

• Faces in the Wild (challenging)

7
Statistics of the Wild Database
8
Measurement

• Confusion Matrix

Models
True Labels
Average Categorization Rate
9
Method

• Features SIFT
• Bag-of-words representation for faces
• Each SIFT feature is considered a codeword
• Build a dictionary based on training samples
• Each face is represented as a histogram over
  codewords
• Learning Naïve Bayes Classifiers

10
Progress

• Datasets done
• Framework almost done
• Results come soon