ACP based face detection Ramin Marx 1 Mai 2007 1 with support from - - PowerPoint PPT Presentation

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

ACP based face detection

Ramin Marx 1 Mai 2007

1with support from Jean-Marc Bo¨

ı and Bernard Fertil

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Introduction

Situation many domains deal with human faces (video surveillance, identification)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Introduction

Situation many domains deal with human faces (video surveillance, identification) Problems humans in the picture? or not? where?

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Introduction

Situation many domains deal with human faces (video surveillance, identification) Problems humans in the picture? or not? where? Goal algorithm, which locates faces in a given picture

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Input picture

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Output picture

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Approach

◮ calculate the faceness of a region R

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Approach

◮ calculate the faceness of a region R ◮ analyze a training database with a huge number of faces

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Approach

◮ calculate the faceness of a region R ◮ analyze a training database with a huge number of faces ◮ extract the most characteristic features and find out how

many of those features contains R

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Idea

◮ treat each of the face images (size n × m) as vector

v ∈ Rn·m

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Idea

◮ treat each of the face images (size n × m) as vector

v ∈ Rn·m

◮ find relationships between dimensions by calculating the

covariances of the training faces

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Idea

◮ treat each of the face images (size n × m) as vector

v ∈ Rn·m

◮ find relationships between dimensions by calculating the

covariances of the training faces

◮ calculate the eigenvectors of that covariance matrix and sort

them descending according to their eigenvalues

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Idea

◮ treat each of the face images (size n × m) as vector

v ∈ Rn·m

◮ find relationships between dimensions by calculating the

covariances of the training faces

◮ calculate the eigenvectors of that covariance matrix and sort

them descending according to their eigenvalues

◮ new base, in which the i-th base vector contains i-th most

information about the data set

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Idea

◮ treat each of the face images (size n × m) as vector

v ∈ Rn·m

◮ find relationships between dimensions by calculating the

covariances of the training faces

◮ calculate the eigenvectors of that covariance matrix and sort

them descending according to their eigenvalues

◮ new base, in which the i-th base vector contains i-th most

information about the data set

◮ we take the first M base vectors and obtain a hyperplane

H ⊂ Rn·m

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Idea

◮ treat each of the face images (size n × m) as vector

v ∈ Rn·m

◮ find relationships between dimensions by calculating the

covariances of the training faces

◮ calculate the eigenvectors of that covariance matrix and sort

them descending according to their eigenvalues

◮ new base, in which the i-th base vector contains i-th most

information about the data set

◮ we take the first M base vectors and obtain a hyperplane

H ⊂ Rn·m

◮ H is the M-dimensional face space, all face vectors lie very

close to it

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

The faceness-test

◮ face vector

v lies close to H ⇔ their distance is small

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

The faceness-test

◮ face vector

v lies close to H ⇔ their distance is small

◮ but what is the distance between

v and H?

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

The faceness-test

◮ face vector

v lies close to H ⇔ their distance is small

◮ but what is the distance between

v and H?

◮ it is the (Euclidian) distance between

v and its projection

• vH ⊂ H

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

The faceness-test

◮ face vector

v lies close to H ⇔ their distance is small

◮ but what is the distance between

v and H?

◮ it is the (Euclidian) distance between

v and its projection

• vH ⊂ H

◮ problem:

v and vH have different bases

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

The faceness-test

◮ face vector

v lies close to H ⇔ their distance is small

◮ but what is the distance between

v and H?

◮ it is the (Euclidian) distance between

v and its projection

• vH ⊂ H

◮ problem:

v and vH have different bases

◮ solution: transform

vH back to Rn·m and then calculate the distance

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Which parameters are interesting?

We have to analyze what impacts

◮ the number of images in the database M,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Which parameters are interesting?

We have to analyze what impacts

◮ the number of images in the database M, ◮ the number of pricipal components M′ to choose,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Which parameters are interesting?

We have to analyze what impacts

◮ the number of images in the database M, ◮ the number of pricipal components M′ to choose, ◮ the comparison algorithm which test the simiarity between

• riginal and reconstructed image,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Which parameters are interesting?

We have to analyze what impacts

◮ the number of images in the database M, ◮ the number of pricipal components M′ to choose, ◮ the comparison algorithm which test the simiarity between

• riginal and reconstructed image,

◮ and the size of the images in the database have.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Tests

All in all, we want to know how the PCA reacts on

◮ images which contain faces from the database,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Tests

All in all, we want to know how the PCA reacts on

◮ images which contain faces from the database, ◮ images which contain faces not in the database,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software PCA Parameters

Tests

All in all, we want to know how the PCA reacts on

◮ images which contain faces from the database, ◮ images which contain faces not in the database, ◮ and on images which contain no faces at all.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 1

Database Olivetti Research Laboratory face database, which contains 400 pictures of faces (40 individuals, 10 poses)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 1

Database Olivetti Research Laboratory face database, which contains 400 pictures of faces (40 individuals, 10 poses) Test 1 100 faces from the database,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 1

Database Olivetti Research Laboratory face database, which contains 400 pictures of faces (40 individuals, 10 poses) Test 1 100 faces from the database, Test 2 100 faces not in the database

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 1

Database Olivetti Research Laboratory face database, which contains 400 pictures of faces (40 individuals, 10 poses) Test 1 100 faces from the database, Test 2 100 faces not in the database Test 3 30 non-face pictures

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Sample pictures with faces from the database

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Sample pictures from faces not in the database

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Sample pictures from non-faces

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

After having made the Principle Component Analysis of our data set with the 300 face images, we take a look at our new base vectors, the Eigenfaces.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Eigenfaces 1 . . . 10 and 20, 25, 30, 40, . . . , 100, 150, 200, 250 and 300.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Start of experiment

Now, we will

◮ take a sample picture,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Start of experiment

Now, we will

◮ take a sample picture, ◮ project it onto the n-dimensional face space,

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Start of experiment

Now, we will

◮ take a sample picture, ◮ project it onto the n-dimensional face space, ◮ vary n from 1 to 300 and

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Start of experiment

Now, we will

◮ take a sample picture, ◮ project it onto the n-dimensional face space, ◮ vary n from 1 to 300 and ◮ consider the reconstruction error (distance between n and its

projection).

Ramin Marx ACP based face detection

dist(original,projection) for the DB-face (lower curve) the non-DB face (upper curve) and the non-face (middle curve).

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ the more Eigenfaces are used, the better is the reconstruction

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ the more Eigenfaces are used, the better is the reconstruction ◮ face from database can be perfectly reconstructed

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ the more Eigenfaces are used, the better is the reconstruction ◮ face from database can be perfectly reconstructed ◮ non-DB face and non-face can also be reconstructed (but not

perfectly)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ the more Eigenfaces are used, the better is the reconstruction ◮ face from database can be perfectly reconstructed ◮ non-DB face and non-face can also be reconstructed (but not

perfectly)

◮ the non-face can be better reconstructed than the non-DB

face

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ the more Eigenfaces are used, the better is the reconstruction ◮ face from database can be perfectly reconstructed ◮ non-DB face and non-face can also be reconstructed (but not

perfectly)

◮ the non-face can be better reconstructed than the non-DB

face

◮ that is bad - we expected the opposite

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ the more Eigenfaces are used, the better is the reconstruction ◮ face from database can be perfectly reconstructed ◮ non-DB face and non-face can also be reconstructed (but not

perfectly)

◮ the non-face can be better reconstructed than the non-DB

face

◮ that is bad - we expected the opposite ◮ but one image is not representative → repeat with more

Ramin Marx ACP based face detection

100 faces from the DB 100 faces not from DB 30 non-faces Average

Closer look at the average picture

many pictures, but the same result: non-faces images are assigned a higher faceness than non-DB faces

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Possible explanation

◮ why are the non-face images closer to the face space than the

non-DB faces?

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Possible explanation

◮ why are the non-face images closer to the face space than the

non-DB faces?

◮ perhaps non-DB faces got a low faceness value because of

disturbing background?

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Possible explanation

◮ why are the non-face images closer to the face space than the

non-DB faces?

◮ perhaps non-DB faces got a low faceness value because of

disturbing background?

◮ apply a mask!

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Possible explanation

◮ why are the non-face images closer to the face space than the

non-DB faces?

◮ perhaps non-DB faces got a low faceness value because of

disturbing background?

◮ apply a mask!

− →

◮

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 2

◮ we repeat the experiment with exactly the same parameters

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 2

◮ we repeat the experiment with exactly the same parameters ◮ except that we apply a mask :-)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 2

◮ we repeat the experiment with exactly the same parameters ◮ except that we apply a mask :-) ◮ unfortunately, almost exactly the same results

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 2

◮ we repeat the experiment with exactly the same parameters ◮ except that we apply a mask :-) ◮ unfortunately, almost exactly the same results ◮ explanation: because the DB was big enough, the PCA was

able to ’ignore’ those unimportant border pixels (that means: don’t consider them while creating the face space)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 3

Database AT&T Face Database, we use 50 pictures of faces (20 individuals, ca. 3 poses)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 3

Database AT&T Face Database, we use 50 pictures of faces (20 individuals, ca. 3 poses) Difference 1 images are illuminated differently → more robust

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 3

Database AT&T Face Database, we use 50 pictures of faces (20 individuals, ca. 3 poses) Difference 1 images are illuminated differently → more robust Difference 2 resolution is much higher (250 × 300 instead of 32 × 32)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 3

Database AT&T Face Database, we use 50 pictures of faces (20 individuals, ca. 3 poses) Difference 1 images are illuminated differently → more robust Difference 2 resolution is much higher (250 × 300 instead of 32 × 32) Difference 3 none of the faces have a background; only the face is visible

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 3

Database AT&T Face Database, we use 50 pictures of faces (20 individuals, ca. 3 poses) Difference 1 images are illuminated differently → more robust Difference 2 resolution is much higher (250 × 300 instead of 32 × 32) Difference 3 none of the faces have a background; only the face is visible Test 1 30 faces from the database

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 3

Database AT&T Face Database, we use 50 pictures of faces (20 individuals, ca. 3 poses) Difference 1 images are illuminated differently → more robust Difference 2 resolution is much higher (250 × 300 instead of 32 × 32) Difference 3 none of the faces have a background; only the face is visible Test 1 30 faces from the database Test 2 30 faces not in the database

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 3

Database AT&T Face Database, we use 50 pictures of faces (20 individuals, ca. 3 poses) Difference 1 images are illuminated differently → more robust Difference 2 resolution is much higher (250 × 300 instead of 32 × 32) Difference 3 none of the faces have a background; only the face is visible Test 1 30 faces from the database Test 2 30 faces not in the database Test 3 30 non-face pictures

Ramin Marx ACP based face detection

Sample pictures of the AT&T database

Recostruction errors of DB-faces, non-DB-faces and non-faces

Average reconstruction errors

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ 20-25 Eigenfaces are enough to distinguish a face from the

DB from any other structure.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Results

◮ 20-25 Eigenfaces are enough to distinguish a face from the

DB from any other structure.

◮ again problems in distinguishing non-DB-faces from non-faces.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 4

◮ we repeat the experiment with the same parameters

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 4

◮ we repeat the experiment with the same parameters ◮ but we reduce the resolution to (37x44) to find out if the

results of the PCA depend on the dimension of the input data

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software Experiment 1 Experiment 2 Experiment 3 Experiment 4

Experiment 4

◮ we repeat the experiment with the same parameters ◮ but we reduce the resolution to (37x44) to find out if the

results of the PCA depend on the dimension of the input data

◮ as expected: the same result

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Questions and conclusions

◮ Why does the PCA react so much on certain non-face images?

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Questions and conclusions

◮ Why does the PCA react so much on certain non-face images? ◮ How to improve the method to compare original picture with

reconstructed picture? Perhaps there are better methods than calculating the Euclidian distance.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Questions and conclusions

◮ Why does the PCA react so much on certain non-face images? ◮ How to improve the method to compare original picture with

reconstructed picture? Perhaps there are better methods than calculating the Euclidian distance.

◮ As the curves suggest, the PCA seems to be better suited for

face recognition than for face detection, because it can very well distinguish between DB-face and any other image, but not between non-DB-faces and non-faces.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Questions and conclusions

◮ Why does the PCA react so much on certain non-face images? ◮ How to improve the method to compare original picture with

reconstructed picture? Perhaps there are better methods than calculating the Euclidian distance.

◮ As the curves suggest, the PCA seems to be better suited for

face recognition than for face detection, because it can very well distinguish between DB-face and any other image, but not between non-DB-faces and non-faces.

◮ Although histogram stretching was used, could more

sophisticated preprocessing improve the results?

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Questions and conclusions

◮ Why does the PCA react so much on certain non-face images? ◮ How to improve the method to compare original picture with

reconstructed picture? Perhaps there are better methods than calculating the Euclidian distance.

◮ As the curves suggest, the PCA seems to be better suited for

face recognition than for face detection, because it can very well distinguish between DB-face and any other image, but not between non-DB-faces and non-faces.

◮ Although histogram stretching was used, could more

sophisticated preprocessing improve the results?

◮ If the above problems were all solved, which heuristics could

be used to increase the speed?

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Face Detection Program

◮ Designed to find faces of any size.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Face Detection Program

◮ Designed to find faces of any size. ◮ Masks can be applied.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Face Detection Program

◮ Designed to find faces of any size. ◮ Masks can be applied. ◮ Different face databases can be easily used.

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Face Detection Program

◮ Designed to find faces of any size. ◮ Masks can be applied. ◮ Different face databases can be easily used. ◮ Heuristics are used which increase the speed about a factor of

20000 (search for regions which are similar to the structure of human eyes)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Face Detection Program

◮ Designed to find faces of any size. ◮ Masks can be applied. ◮ Different face databases can be easily used. ◮ Heuristics are used which increase the speed about a factor of

20000 (search for regions which are similar to the structure of human eyes)

◮ . . .

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Face Detection Program

◮ Designed to find faces of any size. ◮ Masks can be applied. ◮ Different face databases can be easily used. ◮ Heuristics are used which increase the speed about a factor of

20000 (search for regions which are similar to the structure of human eyes)

◮ . . . ◮ but because of the results obtained above, too many non-face

regions are classified as faces and that’s why the program is useless at the moment

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Used papers/software

◮ GNU Scientific Library for matrix calculations like finding

eigenvectors

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Used papers/software

◮ GNU Scientific Library for matrix calculations like finding

eigenvectors

◮ M. Turk and A. Pentland (1991). ’Face recognition using

eigenfaces’. Proc. IEEE Conference on Computer Vision and Pattern Recognition: 586-591. (PDF file available)

Ramin Marx ACP based face detection

Introduction Analysis Experiments Questions and conclusions Face Detection Program Used Papers/Software

Used papers/software

◮ GNU Scientific Library for matrix calculations like finding

eigenvectors

◮ M. Turk and A. Pentland (1991). ’Face recognition using

eigenfaces’. Proc. IEEE Conference on Computer Vision and Pattern Recognition: 586-591. (PDF file available)

◮ http://en.wikipedia.org/wiki/Eigenface and many other

web-pages on PCA, Eigenfaces and face detection

Ramin Marx ACP based face detection