Age Estimation Using Expectation of Label Distribution Learning - - PowerPoint PPT Presentation

▶
age estimation using expectation of label distribution
SMART_READER_LITE
LIVE PREVIEW

Age Estimation Using Expectation of Label Distribution Learning - - PowerPoint PPT Presentation

Jun 16, 2023 234 likes •458 views

Age Estimation Using Expectation of Label Distribution Learning Bin-Bin Gao 1 , Hong-Yu Zhou 1 , Jianxin Wu 1 , Xin Geng 2 1 LAMDA Group, Nanjing University, China 2 PALM Group, Southeast University, China Jul. 19, 2018 Stockholm Background

slide-1
SLIDE 1

Age Estimation Using Expectation

  • f Label Distribution Learning

Bin-Bin Gao1, Hong-Yu Zhou1, Jianxin Wu1, Xin Geng2

1LAMDA Group, Nanjing University, China 2PALM Group, Southeast University, China

  • Jul. 19, 2018 Stockholm
slide-2
SLIDE 2

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

2

Background

  • Identity
  • Emotion
  • Ethnicity
  • Gender
  • Attractiveness
  • Age
  • ……

Face information

This information plays a significant role during face-to-face communication between humans.

slide-3
SLIDE 3

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

3

Background

What is facial age estimation?

It attempts to automatically predict age based on an individual face.

Age Model

Training images: Testing face

Age=37 (years)

slide-4
SLIDE 4

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

4

Background

Potential applications

Law enforcement Security control Recommendations …… Automatic age estimation from face images is an attractive yet challenging topic.

slide-5
SLIDE 5

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

5

Background

200000 400000 600000 800000 1000000 1200000 1400000 1400000

55134 3612 7591

Insufficiency 36 37

Challenges

Imbalance Fine-grained Recognition

slide-6
SLIDE 6

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

6

Related Works

  • MR: Metric Regression [Ranjan et al., FG 2017]
  • DEX: Classification [Rothe et al., IJCV 2016]
  • Ranking [Chen et al., CVPR 2017]
  • DLDL [Gao et al., TIP 2017]

Regression Classification Ranking DLDL

Plenty of deep methods are proposed,

slide-7
SLIDE 7

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

7

Motivation

  • Classification and regression may lead

to an unstable training procedure.

  • There is an inconsistency between

the training objectives and evaluation metric in DLDL and Ranking.

  • Almost all state-of-the-arts have huge

computational cost and storage

  • verhead.

Pervious works have some notable drawbacks,

500M 1G 2G 3G

VGG16 iPhone6 iPhone7 iPhonex

Objective: Fit Distribution Evaluation: MAE

ቐ

slide-8
SLIDE 8

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

8

Proposed Method

Ranking is learning label distribution

Label Distribution Ranking Encoding 50-year-old Normal Distribution c.d.f

slide-9
SLIDE 9

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

9

Proposed Method

Ranking is learning label distribution

There is a linear relationship.

‐ Label distribution can represent more meaningful age information. ‐ Label distribution learning is more efficient. Label Distribution Ranking Encoding CDF

slide-10
SLIDE 10

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

10

Proposed Method

DLDL-v2

 Label Distribution Module

‐ Linear transformation ‐ Label distribution ‐ Loss: KL-Div CNN feature Softmax Label Dis Pred Dis

slide-11
SLIDE 11

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

11

Proposed Method

DLDL-v2

 Expectation Regression Module

This module does not introduce any new parameter. ‐ Expectation layer ‐ Loss: 𝑚1 Label Set

slide-12
SLIDE 12

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

12

Proposed Method

BN BN BN BN BN BN BN BN BN BN BN BN BN

Max Avg

DLDL-v2

 Network Architecture

slide-13
SLIDE 13

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

13

Proposed Method

DLDL-v2

 Jointly Learning (SGD algorithm)

Weight

slide-14
SLIDE 14

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

14

Experiments

Datasets

‐ Apparent age

  • ChaLearn15 (2476+1136)
  • ChaLearn16 (5613+1978)

‐ Real age

  • Morph (55134: 80%+20%)

Evaluation metric

  • MAE：mean average error
  • e-error: It is defined by the ChaLearn.
slide-15
SLIDE 15

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

15

Experiments

Table 1: Comparisons with state-of-the-art methods for apparent and real age estimation. Table 2: Comparisons of model parameters and forward times with state-of-the-arts.

Comparisons with state-of-the-arts

32 images in ms

  • n one M40 GPU.

150× 36× 5.5× 2.6×

slide-16
SLIDE 16

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

16

Experiments

Visual assessment

Good examples Poor examples

slide-17
SLIDE 17

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

17

Experiments

Table 3: Comparison of different methods.

Ablation study

‐ Comparisons

It means that erasing the inconsistency between training and evaluation stages can help us make a better prediction.

slide-18
SLIDE 18

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

18

Experiments

Table 4: The influences of hyper-parameters.

‐ Sensitivity of hyper-parameters

Ablation study

Our method is not sensitive to these hyper-parameters. : Loss weight The number of discrete labels

slide-19
SLIDE 19

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

19

Understanding DLDL-v2

How does DLDL-v2 estimate facial age?

infants adults senior people

The network uses different patterns to estimate different age.

slide-20
SLIDE 20

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

20

Conclusion

 We provide the first analysis and show that the ranking method is in fact learning label distribution implicitly. This result thus unifies existing state-of-the-art facial age estimation methods into the DLDL framework.  We propose an end-to-end learning framework which jointly learns age distribution and regresses single-value age in both feature learning and classifier learning.  We create new state-of-the-art results on facial age estimation tasks using single and small model without external age labeled data or multi-model ensemble.

slide-21
SLIDE 21

http://lamda.nju.edu.cn

http://lamda.nju.edu.cn/gaobb Age Estimation Using Expectation of Label Distribution Learning (IJCAI 2018)

Thanks！

Projects