SCIENCE CHINA Information Sciences, Volume 63, Issue 2: 120102(2020) https://doi.org/10.1007/s11432-019-2733-4

## Ordinal distribution regression for gait-based age estimation

• AcceptedNov 12, 2019
• PublishedJan 15, 2020
Share
Rating

### Abstract

Computer vision researchers prefer to estimate age from face images because facial features provide useful information. However, estimating age from face images becomes challenging when people are distant from the camera or occluded. A person's gait is a unique biometric feature that can be perceived efficiently even at a distance. Thus, gait can be used to predict age when face images are not available. However, existing gait-based classification or regression methods ignore the ordinal relationship of different ages, which is an important clue for age estimation. This paper proposes an ordinal distribution regression with a global and local convolutional neural network for gait-based age estimation. Specifically, we decompose gait-based age regression into a series of binary classifications to incorporate the ordinal age information. Then, an ordinal distribution loss is proposed to consider the inner relationships among these classifications by penalizing the distribution discrepancy between the estimated value and the ground truth. In addition, our neural network comprises a global and three local sub-networks, and thus, is capable of learning the global structure and local details from the head, body, and feet. Experimental results indicate that the proposed approach outperforms state-of-the-art gait-based age estimation methods on the OULP-Age dataset.

### Acknowledgment

This work was supported in part by National Natural Science Foundation of China (Grant No. 61673118), Shanghai Municipal Science and Technology Major Project (Grant No. 2018SHZDZX01), ZJLab, and Shanghai Pujiang Program (Grant No. 16PJD009). We are grateful to the reviewers and the Associate Editor for their constructive comments.

### References

[1] Chang K-Y, Chen C-S, Hung Y-P. Ordinal hyperplanes ranker with cost sensitivities for age estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2011. 585--592. Google Scholar

[2] Chen S, Zhang C, Dong M, et al. Using ranking-CNN for age estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. Google Scholar

[3] Niu Z, Zhou M, Wang L, et al. Ordinal regression with multiple output CNN for age estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 4920--4928. Google Scholar

[4] Pan H, Han H, Shan S, et al. Mean-variance loss for deep age estimation from a face. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 5285--5294. Google Scholar

[5] Shen W, Guo Y, Wang Y, et al. Deep regression forests for age estimation. 2017,. arXiv Google Scholar

[6] Ju Han , Bir Bhanu . Individual recognition using gait energy image.. IEEE Trans Pattern Anal Mach Intell, 2006, 28: 316-322 CrossRef PubMed Google Scholar

[7] Lu J, Tan Y P. Gait-Based Human Age Estimation. IEEE TransInformForensic Secur, 2010, 5: 761-770 CrossRef Google Scholar

[8] Makihara Y, Okumura M, Iwama H, et al. Gait-based age estimation using a whole-generation gait database. In: Proceedings of International Conference on Biometrics (ICB), 2011. 1--6. Google Scholar

[9] Lu J, Tan Y-P. Ordinary preserving manifold analysis for human age estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2010. 90--95. Google Scholar

[10] Lu J, Tan Y P. Ordinary Preserving Manifold Analysis for Human Age and Head Pose Estimation. IEEE Trans Human-Mach Syst, 2013, 43: 249-258 CrossRef Google Scholar

[11] Frank E, Hall M. A simple approach to ordinal classification. In: Proceedings of European Conference on Machine Learning, 2001. 145--156. Google Scholar

[12] Li L, Lin H-T. Ordinal regression by extended binary classification. In: Proceedings of Advances in Neural Information Processing Systems, 2007. 865--872. Google Scholar

[13] Hou L, Yu C-P, Samaras D. Squared earth mover's distance-based loss for training deep neural networks. 2016,. arXiv Google Scholar

[14] Xu C, Makihara Y, Ogi G. The OU-ISIR Gait Database comprising the Large Population Dataset with Age and performance evaluation of age estimation. IPSJ T Comput Vis Appl, 2017, 9: 24 CrossRef Google Scholar

[15] Ricanek K, Tesafaye T. Morph: a longitudinal image database of normal adult age-progression. In: Proceedings of International Conference on Automatic Face and Gesture Recognition (FGR), 2006. 341--345. Google Scholar

[16] Lanitis A, Draganova C, Christodoulou C. Comparing different classifiers for automatic age estimation.. IEEE Trans Syst Man Cybern B, 2004, 34: 621-628 CrossRef PubMed Google Scholar

[17] Yang Z, Ai H. Demographic classification with local binary patterns. In: Proceedings of International Conference on Biometrics (ICB), 2007. 464--473. Google Scholar

[18] Fu Y, Huang T S. Human Age Estimation With Regression on Discriminative Aging Manifold. IEEE Trans Multimedia, 2008, 10: 578-584 CrossRef Google Scholar

[19] Xiao B, Yang X, Zha H, et al. Metric learning for regression problems and human age estimation. In: Proceedings of Pacific-Rim Conference on Multimedia, 2009. 88--99. Google Scholar

[20] Chang K Y, Chen C S. A learning framework for age rank estimation based on face images with scattering transform.. IEEE Trans Image Process, 2015, 24: 785-798 CrossRef PubMed Google Scholar

[21] Rasmussen C E. Gaussian processes in machine learning. In: Proceedings of Advanced Lectures on Machine Learning, 2004. 63--71. Google Scholar

[22] Wada T, Matsumura Y, Maeda S, et al. Gaussian process regression with dynamic active set and its application to anomaly detection. In: Proceedings of the International Conference on Data Mining, 2013. Google Scholar

[23] Makihara Y, Kimura T, Okura F, et al. Gait collector: an automatic gait data collection system in conjunction with an experience-based long-run exhibition. In: Proceedings of International Conference on Biometrics (ICB), 2016. 1--8. Google Scholar

[24] Li X, Makihara Y, Xu C. Gait-based human age estimation using age group-dependent manifold learning and regression. Multimed Tools Appl, 2018, 77: 28333-28354 CrossRef Google Scholar

[25] Crammer K, Singer Y. Pranking with ranking. In: Proceedings of Advances in Neural Information Processing Systems, 2002. 641--647. Google Scholar

[26] Herbrich R, Graepel T, and Obermayer K. Support vector learning for ordinal regression. In: Proceedings of ICANN, 1999. 97--102. Google Scholar

[27] Shashua A, Levin A. Ranking with large margin principle: two approaches. In: Proceedings of Advances in Neural Information Processing Systems, 2003. 961--968. Google Scholar

[28] Maas A L, Hannun A Y, Ng A Y. Rectifier nonlinearities improve neural network acoustic models. In: Proceedings of the 30th International Conference on Machine Learning, Atlanta, 2013. 1--6. Google Scholar

[29] Srivastava N, Hinton G, Krizhevsky A, et al. Dropout: a simple way to prevent neural networks from overfitting. J Machine Learn Res, 2014, 15: 1929--1958. Google Scholar

[30] Sarkar S, Phillips P J, Liu Z. The humanID gait challenge problem: data sets, performance, and analysis.. IEEE Trans Pattern Anal Mach Intell, 2005, 27: 162-177 CrossRef PubMed Google Scholar

[31] Shen W, Guo Y, Wang Y, et al. Deep regression forests for age estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2304--2313. Google Scholar

[32] Parkhi O M, Vedaldi A, Zisserman A, et al. Deep face recognition. In: Proceedings of British Machine Vision Conference (BMVC), 2015. 1: 6. Google Scholar

[33] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. 2014,. arXiv Google Scholar

[34] Kinga D, Adam J B. A method for stochastic optimization. In: Proceedings of International Conference on Learning Representations, 2015. Google Scholar

[35] Russakovsky O, Deng J, Su H. ImageNet Large Scale Visual Recognition Challenge. Int J Comput Vis, 2015, 115: 211-252 CrossRef Google Scholar

[36] Smola A J, Sch?lkopf B. A tutorial on support vector regression. Stat Computing, 2004, 14: 199-222 CrossRef Google Scholar

[37] van der Maaten L, Hinton G. Visualizing data using $t$-SNE J Machine Learn Res, 2008, 9: 2579--2605. Google Scholar

[38] Rothe R, Timofte R, Van Gool L. Deep Expectation of Real and Apparent Age from a Single Image Without Facial Landmarks. Int J Comput Vis, 2018, 126: 144-157 CrossRef Google Scholar

[39] Shen W, Zhao K, Guo Y L, et al. Label distribution learning forests. In: Proceedings of Advances in Neural Information Processing Systems, 2017. 834--843. Google Scholar

[40] Tan Z, Wan J, Lei Z. Efficient Group-n Encoding and Decoding for Facial Age Estimation.. IEEE Trans Pattern Anal Mach Intell, 2018, 40: 2610-2623 CrossRef PubMed Google Scholar

[41] Li W H, Lu J W, Feng J J, et al. Bridgenet: a continuity-aware probabilistic network for age estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1145--1154. Google Scholar

[42] He Y, Zhang J, Shan H. Multi-Task GANs for View-Specific Feature Learning in Gait Recognition. IEEE TransInformForensic Secur, 2019, 14: 102-113 CrossRef Google Scholar

[43] Cai S, Wu Z. Camera calibration with coplanar conics: a unified explanation and ambiguity analysis. IPSJ T Comput Vis Appl, 2018, 10: 14 CrossRef Google Scholar

[44] Wu Z, Huang Y, Wang L. A Comprehensive Study on Cross-View Gait Based Human Identification with Deep CNNs.. IEEE Trans Pattern Anal Mach Intell, 2017, 39: 209-226 CrossRef PubMed Google Scholar

• Figure 1

(Color online) GEIs of subjects of different age and gender in the OULP-Age dataset. The number above each GEI indicates the corresponding age of the subject.

• Figure 2

(Color online) Predictive probability of the $k$-th classifier should not be greater than that of the ($k-1$)-th classifier on an ordinal distribution. Both A and B have the same cross-entropy loss; however, B is preferable to A on an ordinal distribution.

• Figure 3

(Color online) Structure of the proposed ODR-GLCNN. The output layer contains $K-1$ binary classifications incorporating the ordinal information into the current end-to-end learning process.

• Figure 4

(Color online) Age and gender distribution for the OULP-Age dataset. There are 63846 well-labeled GEIs (31093 males and 32753 females, age 2 to 90 years).

• Figure 5

(Color online) Comparisons of age estimation CSs by the proposed approach and state-of-the-art methods on the OULP-Age dataset.

• Figure 6

(Color online) Examples of gait-based age estimation results by using the proposed approach on the OULP-Age dataset. The top three rows show nine successful age estimation examples (MAE smaller than 5 years) for young, middle-aged, and old subjects, respectively. The last row shows nine failure cases (MAE larger than 20 years). The numbers above each image show the subject's ground-truth age and estimated age, respectively.

• Figure 7

(Color online) Feature visualization of CNN (a) and GL-CNN (b). Network features are reduced from 1024 dimensions to 2 dimensions by a $t$-SNE technique. Ages are divided into nine groups. Different colors represent different age groups.

• Figure 8

• Table 1   Comparison of the age estimation MAEs with the proposed approach and state-of-the-art methods on the OULP-Age dataset
 Method MAE CS ($k=5$) (%) SVR [36] 7.66 41.40 MLG [7] 10.98 43.40 OPLDA [9] 8.45 36.50 OPMFA [9] 9.08 34.70 GPR [8] 7.30 43.60 ASSOLPP [24] 6.78 53.00 VGG16 + Mean-Variance [4] 5.59 60.46 ODR-GLCNN (ours) 5.12 66.95
• Table 2   MAE results on the two subsets of the USF dataset using different methods
 Method Gallery Probe A MAE CS ($k=5$) (%) MAE CS ($k=5$) (%) SVR [36] 8.21 37.50 7.83 41.70 MLG [7] 9.45 32.80 9.06 34.40 OPLDA [9] 7.05 43.70 6.76 51.20 OPMFA [9] 6.95 47.30 6.62 52.00 ASSOLPP [24] 6.81 50.50 6.48 52.70 VGG16 + Mean-Variance [4] 6.10 54.20 5.93 56.30 ODR-GLCNN (ours) 5.91 56.60 5.75 59.40
• Table 3   Comparisons among different CNN-based methods on the OULP-Age dataset$^{\rm~a)}$
 Network MAE CS ($k=5$) (%) Time (ms) CNN 5.45 64.64 7.27 $\times~10^{-2}$ VGG16 5.63 63.92 21.9 $\times~10^{-2}$ GL-CNN (ours) 5.24 65.96 8.99 $\times~10^{-2}$
• Table 4   Comparison of different losses with the proposed GL-CNN on the OULP-Age dataset
 Loss MAE CS ($k=5$) (%) Euclidean 6.73 52.95 MAE 6.65 55.16 Squared EMD[13] 6.39 58.34 Cross-Entropy 5.24 65.96 ODL (ours) 5.12 66.95
• Table 5   Influence of human gender for gait-based age estimation in terms of age MAE and gender accuracy
 Method w/o gender w/ gender Accuracy (%) CNN + Euclidean 6.96 6.82 96.70 CNN + Cross-Entropy 5.40 5.34 97.20 VGG16 + Mean-Variance 5.59 5.52 96.70 ODR-GLCNN (ours) 5.12 5.06 97.80
• Table 6   Comparison between proposed approach and state-of-the-art methods on the MORPH Album II dataset in terms of MAE and CS values$^{\rm~a)}$
 Method MAE CS ($k=5$) (%) Protocol OR-CNN [3] 3.27 73.0* RS DEX [38] 3.25 – RS Ranking-CNN [2] 2.96 85.0* RS VGG16 + Mean-Variance [4] 2.41 90.0* RS AGEn [40] 2.93 – RS dLDLF [39] 2.24 – RS DRFs [31] 2.17 91.3 RS BridgeNet [41] 2.38 – RS VGG16 + ODL ($\lambda=0$) 2.30 91.1 RS VGG16 + ODL ($\lambda=0.25$) (ours) 2.16 92.9 RS

a

• #### 0

Citations

• Altmetric

Copyright 2020 Science China Press Co., Ltd. 《中国科学》杂志社有限责任公司 版权所有