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

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 http://lamda.nju.edu.cn Face information • Identity • Emotion • Ethnicity • Gender • Attractiveness • Age • …… This information plays a significant role during face-to-face communication between humans. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 2 Distribution Learning (IJCAI 2018)

Background http://lamda.nju.edu.cn What is facial age estimation? It attempts to automatically predict age based on an individual face. Age=37 Age Model (years) Testing face 3 Training images: Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb Distribution Learning (IJCAI 2018)

Background http://lamda.nju.edu.cn Potential applications Law enforcement Security control Recommendations …… Automatic age estimation from face images is an attractive yet challenging topic. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 4 Distribution Learning (IJCAI 2018)

Background http://lamda.nju.edu.cn Challenges Fine-grained Imbalance Insufficiency Recognition 1400000 1400000 1200000 1000000 800000 600000 400000 55134 3612 7591 200000 0 36 37 Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 5 Distribution Learning (IJCAI 2018)

Related Works http://lamda.nju.edu.cn Plenty of deep methods are proposed, • 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 Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 6 Distribution Learning (IJCAI 2018)

Motivation http://lamda.nju.edu.cn Pervious works have some notable drawbacks, Classification and regression may lead • to an unstable training procedure. There is an inconsistency between • Objective: Fit Dis tribution the training objectives and evaluation ቐ metric in DLDL and Ranking. Evaluation: MAE VGG16 iPhone6 iPhone7 iPhonex Almost all state-of-the-arts have huge • computational cost and storage overhead . 500M 1G 2G 3G Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 7 Distribution Learning (IJCAI 2018)

Proposed Method http://lamda.nju.edu.cn Ranking is learning label distribution Ranking Encoding Label Distribution c.d.f Normal Distribution 50-year-old Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 8 Distribution Learning (IJCAI 2018)

Proposed Method http://lamda.nju.edu.cn Ranking is learning label distribution Ranking Encoding Label Distribution CDF There is a linear relationship. ‐ Label distribution can represent more meaningful age information. ‐ Label distribution learning is more efficient. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 9 Distribution Learning (IJCAI 2018)

Proposed Method http://lamda.nju.edu.cn DLDL-v2  Label Distribution Module ‐ Linear transformation CNN feature ‐ Label distribution Softmax ‐ Loss: KL-Div Label Dis Pred Dis Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 10 Distribution Learning (IJCAI 2018)

Proposed Method http://lamda.nju.edu.cn DLDL-v2  Expectation Regression Module ‐ Expectation layer Label Set ‐ Loss: 𝑚 1 This module does not introduce any new parameter. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 11 Distribution Learning (IJCAI 2018)

Proposed Method http://lamda.nju.edu.cn DLDL-v2  Network Architecture Max BN BN BN BN BN BN BN BN BN Avg BN BN BN BN Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 12 Distribution Learning (IJCAI 2018)

Proposed Method http://lamda.nju.edu.cn DLDL-v2  Jointly Learning (SGD algorithm) Weight Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 13 Distribution Learning (IJCAI 2018)

Experiments http://lamda.nju.edu.cn 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. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 14 Distribution Learning (IJCAI 2018)

Experiments http://lamda.nju.edu.cn Comparisons with state-of-the-arts 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. 32 images in ms on one M40 GPU. 150 × 36 × 5.5 × 2.6 × Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 15 Distribution Learning (IJCAI 2018)

Experiments http://lamda.nju.edu.cn Visual assessment Good examples Poor examples Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 16 Distribution Learning (IJCAI 2018)

Experiments http://lamda.nju.edu.cn Ablation study ‐ Comparisons Table 3: Comparison of different methods. It means that erasing the inconsistency between training and evaluation stages can help us make a better prediction. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 17 Distribution Learning (IJCAI 2018)

Experiments http://lamda.nju.edu.cn Ablation study ‐ Sensitivity of hyper-parameters Table 4: The influences of hyper-parameters. : Loss weight The number of discrete labels Our method is not sensitive to these hyper-parameters . Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 18 Distribution Learning (IJCAI 2018)

Understanding DLDL-v2 http://lamda.nju.edu.cn How does DLDL-v2 estimate facial age? infants adults senior people The network uses different patterns to estimate different age. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 19 Distribution Learning (IJCAI 2018)

Conclusion http://lamda.nju.edu.cn  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. Age Estimation Using Expectation of Label http://lamda.nju.edu.cn/gaobb 20 Distribution Learning (IJCAI 2018)

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