Viola and Jones Object Detector Ruxandra Paun EE/CS/CNS 148 - Presentation 04.28.2005.

Viola and Jones Object Detector

Ruxandra Paun EE/CS/CNS 148 - Presentation

04.28.2005

Fast! 15 times faster than any previous

approach 384 by 288 pixel images detected at

15 frames per second on a conventional

700 MHz Intel Pentium III

Robust Real-Time Face Detection 3 key contributors:

- a new image representation: the “Integral Image” - a simple and effective classifier, based on the AdaBoost learning algorithm - combining the classifiers in a

“cascade”

Detection basis: Features

Integral Image

Computing features

Classifier: using AdaBoost 160,000 features for every sub-window Very small number of these features

can be combined to form an effective classifier

AdaBoost: constrain each week classifier to depend on a single feature

each stage of boosting = new week classifier selection = feature selection

First and Second Features Selected by AdaBoost

ROC curve for a 200 feature classifier

The Cascade combining successively more

complex classifiers in a cascade structure

38 stages

ROC curves: cascaded vs. monolithic classifier

-> not significantly different accuracy

-> but the cascade class. almost 10 times faster

Results

Training dataset: 4916 images

ROC Curves for Face Detection

Comparing Viola-Jones with Other Systems

More: Detecting Walking Pedestrians

Integrating image intensity with motion information Efficient, detects pedestrians at small

scales, and has a very low false positive rate

Works on low resolution images and under difficult weather conditions (rain, snow)

Extracting motion information

Training Set Samples

QuickTime™ and aYUV420 codec decompressor

are needed to see this picture.

Questions?