Overview of Results

Tried at least two methods for most of the features

 

Good news!

Technique
Feature
a) Maximum Likelihood(ML)
Sex
b) Neural Networks (NN)
Age
c) Generalized Linear Discriminant(GLD)
Expression
d) Multi-Class GLD
Race
e)Support Vector Machine (SVM)
Bandana

synthetic faces made the difference

Bad news

Technique
Con
a
Not everything that "looks" Gaussian is likely to be distributed as Gaussian.
b
 Performs poorly on classes for which training data is small compared to the data from the other classes.

 

Feature

 

class i/class j/...

 
Best result (% class i/%class j/...)

Sex

(male/female)
81% / 66% ML

Expression

(smiling/serious/funny)
75% / 88% / 5% NN

Expression

(smiling/serious)
83% / 81% GLD

Age

(child/teen/adult/senior)
50%/22%/83%/44% NN

Race

(Asian/Black/Hispanic/White)
69%/47%/0%/85% multi-GLD

Hat

(with/without)
75% / 93% SVM

Bandana

(with/without)
38% / 98% GLD

Glasses
(with/without)
25% / 97% GLD

Moustache

(with/without)
80% / 95% GLD

Beard

(with/without)
93% / 87% GLD

Moustache-Beard
(with/without)
55% / 99% GLD

 
Back
Contents
Next