K E C H E N 1 , S H A O G A N G G O N G 1 , T A O X I A N G 1 , C H E N C H A N G E L O Y 2
1 . Q U E E N M A R Y , U N I V E R S I T Y O F L O N D O N2 . T H E C H I N E S E U N I V E R S I T Y O F H O N G K O N G
CUMULATIVE ATTRIBUTE SPACE FOR AGE AND CROWD DENSITY ESTIMATION
CVPR 2013, Portland, Oregon
PROBLEMS
How old are they?
How many persons are in the scene?
What is the head pose (viewing angles) of this person?
A REGRESSION FORMULATIONOriginal images/frames
Facial images
Crowd frames
AAM feature
Segment feature
Edge feature
Texture feature
Feature extraction
Feature space Label space
LabelsLearning the mapping
Regression
CHALLENGE – FEATURE VARIATION
The same age
Extrinsic conditions: Lighting conditions; Viewing angles Intrinsic conditions: aging process of different people glasses, hairstyle, gender, ethnicity
Feature
CHALLENGE – FEATURE VARIATION
The same person count
Extrinsic conditions: Lighting conditions; Viewing angles Intrinsic conditions: occlusion, density distribution in the scene
Feature
CHALLENGE – SPARSE AND IMBALANCED DATA
Data distribution of FG-NET Dataset
Max number of samples for each age group is 46
CHALLENGE – SPARSE AND IMBALANCED DATA
Data distribution of UCSD Dataset
RELATED WORKS
• Most focused on feature variation challenge
• Few focused on sparse and imbalanced data challenge
• Two challenges are related
1. Improve feature robustness [Guo et al, CVPR, 2009; Guo et al, TIP, 2012; Ryan et al, DICTA, 2009; Zhang et al, IEEE T ITS, 2011].
2. Improve regressor
[Guo et al, TIP 2008; Chang et al, CVPR 2011; Chao et al, PR 2013; Chan et al, CVPR 2008; Chen et al, BMVC 2012]
OUR APPROACHSolution:• Attribute Learning can address data sparsity problem
--Exploits the shared characteristics between classesHas sematic meaningDiscriminative
Problems:• Applied successfully in classification but not in
regression• How to exploit cumulative dependent nature of
labels in regression?…… …… ……
Age 20 Age 21 Age 60
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
LIMITATION OF NON-CUMULATIVE ATTRIBUTE
Age 200
1
…
20th
0…0
Age 6060th0
…
0
0
0
1
…
0…
0
0…0
0
21st0
1
…
0
…
0
0
0
0
Age 21
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
OUR FRAMEWORK
Imagery Features xi
Facial images Crowd frames
Labels yi
Regression Learning
Cumulative Attributes ai
Feature Extraction
Multi-output Regression Learning
Regression Mapping
Conventional frameworks
1 1 0 0… …1 1 2 yi yi+1 N
JOINT ATTRIBUTE LEARNING
• Joint Attribute Learning
with quadratic loss function
• Regression Learning with attribute representation as input is not limited to a specific regression model
min 12‖𝐖‖ 2
𝐹+𝐶∑𝑖=1
N
‖𝐚𝑖𝑇−(𝐱 𝑖
𝑇𝐖+𝐛)‖ 2𝐹
min 12‖𝐰 𝑗‖ 2
2+𝐶∑𝑖=1
N
𝑙𝑜𝑠𝑠(𝑎𝑖𝑗 , 𝑓 𝑗 (𝐱 𝑖))¿¿
COMPARATIVE EVALUATION
Age EstimationCA-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;
COMPARATIVE EVALUATION
Crowd Counting
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;
CUMULATIVE (CA) VS. NON-CUMULATIVE (NCA)
Crowd Counting
Age Estimation
ROBUSTNESS AGAINST SPARSE AND IMBALANCED DATA
Age Estimation
Crowd Counting
FEATURE SELECTION BY ATTRIBUTES
Shape plays a more important role than texture when one is younger.
CONCLUSION
• A novel attribute framework for regression
• Exploits cumulative dependent nature of label space
• Effectively addresses sparse and imbalanced data problem
Thanks a lot for your attention! Any questions?
Welcome to our poster 3A-2 for more details.
Ke Chen Shaogang Gong Tao Xiang Chen Change Loy Ph.D student Professor Associate Professor Assistant Professor