sign classification boosted cascade of classifiers using university of southern california thang...

26
Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh [email protected] Eunyoung Kim [email protected] Li Zhang [email protected] Yuping Lin [email protected] Computer Vision group:

Upload: giles-bond

Post on 13-Jan-2016

218 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Sign Classification

Boosted Cascade of Classifiers

using

University of Southern California

Thang Dinh [email protected] Kim [email protected] Zhang [email protected] Lin [email protected]

Computer Vision group:

Page 2: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

2. Problem statement

Content

1. Sign classification: Applications and Challenges

3. Why Boosted Cascade of Classifiers ?

4. Sign detection and classification with Cascade5. Experimental Results

Page 3: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Sign classification: Applications & Challenges

Applications

Sign language

……..

Virtual Reality

Robot Control

Sign Classification

Page 4: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Sign classification: Applications & Challenges

Challenges

1. Detection Rate

2. Computation Time

An interpreting system that interprets wrongly, a robot that always misunderstands commands… can not be employed.

Robots that need 30s to understand a command telling them to do something immediately, Systems that need 30s to interpret each word of speakers… will not be employed either.

need a Fast and Robust system

Page 5: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

2. Problem statement

Content

1. Sign classification: Applications and Challenges

3. Why Boosted Cascade of Classifiers ?

4. Sign detection and classification with Cascade5. Experimental Results

Page 6: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Problem statement

Human Gesture:

• Dynamic Gesture: requires motion of hand, body…• Static Gesture: static pose of hand, body…

• Finger Spelling

(ASL System)

Page 7: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Problem statement

Implemented system:

Challenges:

• Detection rate, Computation time• Many similar signs

….

Page 8: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

2. Problem statement

Content

1. Sign classification: Applications and Challenges

3. Why Boosted Cascade of Classifiers ?

4. Sign detection and classification with Cascade5. Experimental Results

Page 9: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Why Boosted Cascade of Classifiers

• Cascade of boosted classifiers with Haar wavelet features (Viola and Jones) is currently ‘The state of the art’ in face detection.

• Cascade was brought to field of hand detection by Eng-Jon with impressive results.

Page 10: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Why Boosted Cascade of Classifiers

Face detection system of Viola and Jones is:

• 15 times faster than Rowley’s (double layer Neural Network)• 600 times faster than Schneiderman-Kanade’s (Statistics)

Cascade of boosted classifiers seems to be a good approach for the Sign classification problem.

Page 11: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

2. Problem statement

Content

1. Sign classification: Applications and Challenges

3. Why Boosted Cascade of Classifiers ?

4. Sign detection and classification with Cascade5. Experimental Results

Page 12: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Haar wavelet features

• Haar wavelet features (Papageorgiou )

• Feature set was enlarged by Lien Hart to 45º rotated.

• We proposed Double-L for Sign detection

Page 13: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Integral Image – Fast calculation of Haar features

• There are millions of features that need processing and each feature itself needs time to be calculated makes the training process time-consuming

• Viola and Jones proposed Integral Image in order to reduce the feature calculation complexity, which results in less computation time

Definition of intergral image at point (x, y)

Sum(D) = 1 + 4 – (2 + 3)

Then:

Page 14: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Adaboost

• Proposed by Freund and Schapire

• Combine many ‘weak’ hypothesis to form a ‘strong’ one

• ‘weak’ classifier ht only need to be better than chance

Page 15: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Adaboost with Haar wavelet features

Advantages:

• Adaboost adjusts adaptively the errors of the weak hypotheses

• Require large amount of training samples

• Weak Hypothesis learnt from this algorithm can run very fast because of simple calculation of Haar feature, which speeds up the whole system

Disadvantages:

• The training process is rather time-consuming because the algorithm needs to check through millions of features extracted from thousands of samples.

Page 16: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Cascade of classifiers

• Cascade of boosted classifiers is a tree of classifiers where classifier lying at each stage is better than the last• Only those input patterns having passed through all the layers are considered objects

Simple backgrounds can be easily rejected by one-feature classifier

Page 17: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Sign Classifiers

• Each Sign Detectors Di is a cascade of boosted classifier evaluating an input image to give out value vi which is then compared to the threshold of it

• 24 Sign Detectors are combined to form a Sign Classifier

C = signi | Di = signi and g(Di)=max{g(Dj)|j=1, 2,… 24}

Where g(Di) = |vi – i|

Page 18: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Sign Classifiers

Page 19: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

2. Problem statement

Content

1. Sign classification: Applications and Challenges

3. Why Boosted Cascade of Classifiers ?

4. Sign detection and classification with Cascade5. Experimental Results

Page 20: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Experimental Results

Samples Collection

• Collect raw image

• Find the ‘key’ point (which make it different from others)• Cut and Align the image (base on ‘key’ point)

Image B - Raw

Image B - Focus

Image B - Aligned

• Aligned images are finally made greyscale

Page 21: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Experimental Results

Training Process

• First stage:

• Positive: Thousands of aligned sign images• Negative: Background only – Buildings, paintings, trees, other parts of human body… (also non-sign)

• Second stage:

• Positive: also thousands of aligned sign images• Negative: aligned images of other signs

Fast eliminate simple background

Distinguish each sign from others

Page 22: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Experimental Results

Results

1. Detectors

• We have trained 24 detectors for 24 signs• Average number of stages: 14

• Total features: approx 123

• Detection rate: 90% - 100%

• Test Images were divided into 2 groups

• Simple background

• Complex background

DR = 100%

FA = 0%

DR = …

FA = …

Page 23: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Experimental Results

Results

1. Detectors (cont)

Hit rate diagram

Page 24: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Experimental Results

Results

1. Detectors (cont)

False alarm

High FA rate due to similarity between signs

Page 25: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Experimental Results

Results

2. Classifiers

Test Images: 600

Correct rate: 83%

Page 26: Sign Classification Boosted Cascade of Classifiers using University of Southern California Thang Dinh thang.dinh@usc.edu Eunyoung Kim eunyoung.kim@usc.edu

Jan 2007

Thank you

Thank you