Implementation and Validation of Fixed

Download Report

Transcript Implementation and Validation of Fixed

Detecting mouth with an AdaBoost
cascade based on Haar-like features
P.Viola, M.Jones Robust Real-time
Object Detection.
Second International Workshop on
statistical and computational theories
of vision – Modeling, Learning,
computing and sampling. 2001
Santa Rossi
(GET-ENST)
Mardi 15 novembre 2005
1
Advantages using OpenCV Haar object
detection






Face detector already implemented
Its only argument is a xml file
Detection at any scale
Face detection at 15 frames per second for
384*288 pixel images
90% objects detected
10-6 false positive rate (FAR)
Mardi 15 novembre 2005
2
Integral Image

The value of the integral
image at point (x; y) is the
sum of all the pixels above
and to the left.
Integral Image

The sum of the pixels which
lie within the white
rectangles are subtracted
from the sum of pixels in the
grey rectangles.
Example rectangle features
Mardi 15 novembre 2005
3
Adaboost classifier


Selects a small number
of critical visual
features
Combines a collection
of weak classification
functions to form a
strong classifier
Mardi 15 novembre 2005
The first and second features
selected by AdaBoost for face
detection
4
First and second features selected for
mouth detection
Mardi 15 novembre 2005
5
Cascade


Stages in the cascade are constructed by training
classifiers using Adaboost
Training this detector takes weeks but it is done
once and for all. Then, it processes 15 frames per
second
Mardi 15 novembre 2005
6
Experiments

BANCA database


WM + G1 as training
dataset: 3420 mouth
images
Negative examples =



WM + G1 (face – mouth)
images
WM + G1 (bottom part of
the face – mouth)
images
G2 test data set: 3117
images
Mardi 15 novembre 2005
7
Experiments

First experiments:
 2 weeks to train
 16 stages cascades
 First results on training
data

Cascade 1


Cascade 2


Pos=93.71 %
First results on test data

Cascade 1


Pos=94.42 %
Cascade 2


Pos=92.25 %
Pos=95.77 %
Eyes permissive !
Mardi 15 novembre 2005
8
Xml file
<?xml version="1.0"?>

<opencv_storage>

<detect_bouche_demi_face_20_20 type_id="opencv-haar-classifier">

<size>

20 20</size>

<stages>

<_>

<!-- stage 0 -->

<trees>

<_>

<!-- tree 0 -->

<_>

<!-- root node -->

<feature>

<rects>

<_>

0 2 20 12 -1.</_>

<_>

0 5 20 6 2.</_></rects>

<tilted>0</tilted></feature>

<threshold>-0.0563905797898769</threshold>

<left_val>0.9257497191429138</left_val>

<right_val>-0.9348185062408447</right_val></_></_>

<_>
Mardi
15 novembre

<!-- 2005
tree 1 -->

9
Perspectives






Test G2
Add negative samples
Change negative samples process
This detector has a high acceptance rate, we
can add a verification test to validate the
detected object
Implementation on PDA
Learn and test with Qtek database
Mardi 15 novembre 2005
10