Real-Time Detection, Alignment and Recognition of Human Faces

1
Real-Time Detection, Alignment and Recognition of
Human Faces

• Rogerio Schmidt Feris
• Changbo Hu
• Matthew Turk
• Pattern Recognition Project
• June 12, 2003

2
Overview

• Introduction
• FERET Dataset
• Face Detection
• Face Alignment
• Face Recognition
• Conclusions

3
Introduction
Detection
Alignment
Recognition
4
Introduction

• Why this is a difficult problem?
• Facial Expressions, Illumination Changes, Pose,
  etc.
• Assumption Frontal view faces
• Objectives
• Develop a fully automatic system, suitable for
  real-time applications.
• Evaluate it on a large dataset.

5
FERET DataSet

• 1196 different individuals
• Probe Sets
• FB Different facial expressions
• FC Different illumination conditions
• DUP1 Different days
• DUP2 Images taken at least 1 year after

6
Face Detection

• State-of-the-art Learning-based approaches
• Neural Nets Rowley et al, PAMI 98
• SVMs Heisele and Poggio, CVPR 01
• Boosting Viola and Jones, ICCV 01
• Want to know more?
• Detecting Faces in Images a Survey M. Yang,
  PAMI 02

7
Face Detection

• Viola and Jones, 2001
• Simple features, which can be computed very
  fast.
• A variant of Adaboost is used both to select
  the features and to train the classifier.
• Classifiers are combined in a cascade which
  allows background regions of the image to be
  quickly discarded.

8
Face Detection
Time 100ms (PIV 1.6Ghz)
9
Face Alignment

• State-of-the-art Deformable Models
• Bunch-Graph approach Wiskott, PAMI 98
• Active Shape Models Cootes, CVIU 95
• Active Appearance Models Cootes, PAMI 01

10
Face Alignment

• Active Appearance Model (AAM)

Statistical Shape Model (PCA)
Statistical Texture Model (PCA)

• AAM Search

11
Face Alignment

• Problem Partial Occlusion
• Active Wavelet Networks (AWN) (submitted to
  BMVC03)
• Main idea Replace AAM texture model by a
  wavelet network

12
Face Alignment
Similar performance to AAM in images under normal
conditions.
More robust against partial occlusions.
13
Face Alignment

• Using 9 wavelets, the system requires only 3 ms
  per iteration (PIV 1.6Ghz). In general, at most
  10 iterations are sufficient for good convergence.

14
Face Recognition

• State-of-the-art Subspace Techniques
• PCA, FDA, Kernel PCA, Kernel FDA, ICA, etc.
• Want to know more?
• Face Recognition a Literature Survey W. Zhao,
  2000

15
Face Recognition

• www.cs.colostate.edu/evalfacerec/
• Preprocessing
• Line up eyes, histogram equalization, masking
• Subspace Training (PCA)
• Classification (Nearest-neighbor)

16
Face Recognition
17
Face Recognition
18
Face Recognition
19
Face Recognition
20
Conclusions

• An efficient, fully automatic system for face
  recognition was presented and evaluated.
• Future Work
• Alignment multiresolution search
• View-based face recognition
• Explicit illumination model
• Live demo

21
Face Recognition
22
Face Recognition
23
Face Recognition