+ All Categories
Home > Documents > Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2....

Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2....

Date post: 01-Apr-2015
Category:
Upload: darian-varnes
View: 213 times
Download: 0 times
Share this document with a friend
Popular Tags:
24
Ke Chen 1 , Shaogang Gong 1 , Tao Xiang 1 , Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group presentation by Minh Hoai Cumulative Attribute Space for Age and Crowd Density Estimation
Transcript
Page 1: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Ke Chen1, Shaogang Gong1, Tao Xiang1, Chen Change Loy2

1. Queen Mary, University of London2. The Chinese University of Hong Kong

VGG reading group presentation by Minh Hoai

Cumulative Attribute Space for Age and Crowd Density Estimation

Page 2: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Tasks

How old are they?

How many people?

What is the head angle?

Page 3: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

A Regression FrameworkInput images

AAM feature

Segment feature

Edge feature

Texture feature

Feature extraction

Features Label space

Label (age, count)Learn the mapping

Regression

Page 4: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Data distribution of FG-NET Dataset

Challenge – Sparse and Unbalanced data

Page 5: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Data distribution of UCSD Dataset

Challenge – Sparse and Unbalanced data

Page 6: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Proposed Approach

Solution:• Attribute Learning can address data sparsity problem --

Exploits the shared characteristics between classesHas sematic meaning

Question to address:• How to exploit cumulative dependent nature of labels in

regression?

…… …… ……

Age 20 Age 21 Age 60

Page 7: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Cumulative Attribute

Age 20

1

1

0

1

…20

0…0

the rest

Cumulative attribute (dependent)

Vs.

0

1

20th

0…

0

Non-cumulative attribute (independent)

0

0

Page 8: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Limitation of Non-cumulative Attribute

Age 200

1

20th

0

…0

Age 60

60th0

0

0

0

1

0

0

0…0

0

21st

0

1

0

0

0

0

0

Age 21

Page 9: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Age 21

Advantages of Cumulative Attribute

Age 20

1

1

0

1

…20

0

…0

the rest

Age 60

1

1

1

…60

0…

0

1

0

… 1…1 attribute changes

1

1…21

0

0

1

1

0

40 attributes change

Page 10: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Proposed Framework

xiFeature vector (e.g., intensity)

yiLabel

(e.g., age)

1

1

1

… …0

0

aiCumulative

attribute

yi

1

2

The task

Regressor Regressor

Page 11: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Proposed Framework

xiFeature vector (e.g., intensity)

yiLabel

(e.g., age)

1

1

1

… …0

0

aiCumulative

attribute

yi

1

2

Our task

Regressor Regressor

How are these regressors learned?

Can use any regression method: Support Vector Regression, Ridge Regression

See next slide!

Page 12: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Regressor for Cumulative Attributes

Regularization Regression error

# of training dataCumulative

attribute Image feature

vectorParameters

to learn

Closed-form solution:

Page 13: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Experiments

Page 14: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Baseline Methods and Name Abbreviation

xiFeature vector

yiLabel

Cumulative attributes

Non-Cumulative attributes

Support Vector Regression (SVR)

SVR

SVR

CA-SVR

NCA-SVR

…1 1 0 0…1

1 2 yi

…1 0 0…00

1 2 yi

Page 15: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Cumulative (CA) vs. Non-cumulative (NCA)

Age Estimation

Mean absolute error(lower is better)

Percentage of prediction within 5 years(higher is better)

Page 16: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Cumulative (CA) vs. Non-cumulative (NCA)

Crowd Counting

Mean absolute error(lower is better)

Mean squared error(lower is better)

Mean deviation error(lower is better)

Page 17: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Crowd Counting Results

CA-RR: our method; LSSVR: Suykens et al, IJCNN, 2001; KRR: An et al, CVPR, 2007; RFR: Liaw et al, R News, 2002; GPR: Chan et al, CVPR, 2008; RR: Chen et al, BMVC, 2012;

Based on regression

Proposed method, RR: Ridge Regression

Ridge Regression without attributes

Page 18: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Age Estimation Results

CA-SVR: our method; AGES: Geng et al, TPAMI, 2007; RUN: Yan et al, ICCV, 2007; Ranking: Yan et al, ICME, 2007; RED-SVM: Chang et al, ICPR, 2010; LARR: Guo et al, TIP, 2008; MTWGP: Zhang et al, CVPR, 2010; OHRank: Chang et al, CVPR, 2011; SVR: Guo et al, TIP, 2008;

Proposed method, SVR: Support Vector Regression

Not based on regression

What is OHRank?

Page 19: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

OHRank - Ordinal Hyperplanes Ranker

SVM score for older than k

Delta 0/1 function

This is 104 slower than closed-form solution of regression

Page 20: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Robustness Against Sparse and Unbalanced Data

Age Estimation

Crowd Counting

(Effects of removing random/certain label groups)

Page 21: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Feature Selection by Attributes

Shape plays a more important role than texture for younger ages.

Page 22: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Summary

• Has a simple and neat idea • Exploits cumulative dependent nature of label space• Addresses sparse and unbalanced data problem

Page 23: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Support Vector Regression

Page 24: Ke Chen 1, Shaogang Gong 1, Tao Xiang 1, Chen Change Loy 2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong VGG reading group.

Datasets


Recommended