Dynamic Label Graph Matching for Unsupervised Video Re-Identification

Posted by JoselynZhao on October 22, 2019

Abstract

This pa-per focuses on cross-camera label estimation, which can be subsequently used in feature learning to learn robust re-IDmodels.

Specifically, we propose to construct a graph forsamples in each camera, and then graph matching schemeis introduced for cross-camera labeling association.

this paper propose a dynamic graphmatching (DGM) method.

DGM iteratively updates theimage graph and the label estimation process by learninga better feature space with intermediate estimated labels.

DGM is advantageous in two aspects:

1) the accuracy of es-timated labels is improved significantly with the iterations;

2) DGM is robust to noisy initial training data.

Code is available at www.comp.hkbu.edu.hk/%7emangye/

Introduction

Different from theexisting unsupervised person re-ID methods, this paper isbased on a more customized solution, i.e., cross-cameralabel estimation. In other words, we aim to mine the la-bels (matched or unmatched video pairs) across cameras.With the estimated labels, the remaining steps are exactlythe same with supervised learning.

In light of the above discussions, this paper proposes adynamic graph matching (DGM) method to improve thelabel estimation performance for unsupervised video re-ID(the main idea is shown in Fig. 1). Specifically, our pipelineis an iterative process. In each iteration, a bipartite graph isestablished, labels are then estimated, and then a discrim-inative metric is learnt. Throughout this procedure, labelsgradually become more accurate, and the learnt metric morediscriminative. Additionally, our method includes a labelre-weighting strategy which provides soft labels instead ofhard labels, a beneficial step against the noisy intermediatelabel estimation output from graph matching. 在这里插入图片描述

the main contributions are summarized as follows:

We propose a dynamic graph matching (DGM) methodto estimate cross-camera labels for unsupervised re-ID, which is robust to distractors and noisy initial train-ing data.

Our experiment confirms that DGM is only slightly in-ferior to its supervised baselines and yields competi-tive re-ID accuracy compared with existing unsuper-vised re-ID methods on three video benchmarks.

Methods

Graph Matching for Video Re-ID

在这里插入图片描述 在这里插入图片描述 在这里插入图片描述 在这里插入图片描述

在这里插入图片描述 在这里插入图片描述 关于公示5的解释: 在这里插入图片描述 在这里插入图片描述 这张图告诉了我们,video based和image based的区别。video 对应多张图,而image对应一张图.

Dynamic Graph Matching

However, there still remains two obvi-ous shortcomings:

we need to learna discriminative feature space to optimize the graphmatching results.

Therefore, it is reasonable to re-encode theweights of labels for overall learning, especially forthe uncertain estimated positive video pairs. 在这里插入图片描述

Label Re-weighting

在这里插入图片描述

在这里插入图片描述 在这里插入图片描述 在这里插入图片描述 The label re-weighting scheme has the following advan-tages:

(1) for positive video pairs, it could filter some falsepositives and then assign different positive sample pairs dif-ferent weights;

(2) for negative video pairs, a number ofeasy negatives would be filtered. The re-weighing schemeis simple but effective as shown in the experiments.

Metric Learning with Re-weighted Labels

在这里插入图片描述

在这里插入图片描述

Iterative Updating

The proposed approach is summarized in Algorithm 1. 在这里插入图片描述

References

[1] A. Beck and M. Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAMjournal on imaging sciences, 2009. 5 [2] M. Cho, J. Lee, and K. M. Lee. Reweighted random walksfor graph matching. In ECCV, 2010. 2 [3] H. Fan, L. Zheng, and Y. Yang. Unsupervised person re-identification: Clustering and fine-tuning. ar Xiv preprintar Xiv:1705.10444, 2017. 2 [4] S. Hamid Rezatofighi, A. Milan, Z. Zhang, Q. Shi, A. Dick,and I. Reid. Joint probabilistic data association revisited. InICCV, 2015. 2, 3 [5] S. Hamid Rezatofighi, A. Milan, Z. Zhang, Q. Shi, A. Dick,and I. Reid. Joint probabilistic matching using m-best solu-tions. In CVPR, 2016. 2, 3 [6] M. Hirzer, C. Beleznai, P. M. Roth, and H. Bischof. Personre-identification by descriptive and discriminative classifica-tion. In Image analysis, 2011. 5 [7] S. Karanam, Y. Li, and R. J. Radke. Person re-identificationwith discriminatively trained viewpoint invariant dictionar-ies. In ICCV, 2015. 7, 8 [8] F. M. Khan and F. Bremond. Unsupervised data associationfor metric learning in the context of multi-shot person re-identification. In AVSS, 2016. 2, 7, 8 [9] E. Kodirov, T. Xiang, Z. Fu, and S. Gong. Person re-identification by unsupervised l1 graph learning. In ECCV,2016. 1, 2, 7, 8[10] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenetclassification with deep convolutional neural networks. InNIPS, 2012. 7 [11] X. Lan, A. J. Ma, P. C. Yuen, and R. Chellappa. Joint sparserepresentation and robust feature-level fusion for multi-cuevisual tracking. IEEE TIP, 2015. 1 [12] X. Lan, P. C. Yuen, and R. Chellappa. Robust mil-basedfeature template learning for object tracking. In AAAI, 2017.1 [13] M. Leordeanu, R. Sukthankar, and M. Hebert. Unsupervisedlearning for graph matching. In IJCV, 2012. 2, 3 [14] S. Liao, Y. Hu, X. Zhu, and S. Z. Li. Person re-identificationby local maximal occurrence representation and metriclearning. In CVPR, 2015. 1, 6, 7, 8 [15] S. Liao and S. Z. Li. Efficient psd constrained asymmetricmetric learning for person re-identification. In ICCV, 2015.5, 6, 7, 8 [16] K. Liu, B. Ma, W. Zhang, and R. Huang. A spatio-temporal appearance representation for viceo-based pedes-trian re-identification. In ICCV, 2015. 1, 3, 5, 7, 8 [17] Z. Liu, D. Wang, L. Zheng, and H. Lu. A labeling-by-searchapproach for unsupervised person re-identification. In ICCV,2017. 2 [18] A. J. Ma, P. C. Yuen, and J. Li. Domain transfer support vec-tor ranking for person re-identification without target cameralabel information. In ICCV, 2013. 2 [19] T. Matsukawa, T. Okabe, E. Suzuki, and Y. Sato. Hierarchi-cal gaussian descriptor for person re-identification. In CVPR,2016. 1 [20] N. Mc Laughlin, J. Martinez del Rincon, and P. Miller. Re-current convolutional network for video-based person re-identification. In CVPR, 2016. 1 [21] R. Panda, A. Bhuiyan, V. Murino, and A. K. Roy-Chowdhury. Unsupervised adaptive re-identification in openworld dynamic camera networks. In CVPR, 2017. 2 [22] P. Peng, T. Xiang, Y. Wang, and a. et. Unsupervised cross-dataset transfer learning for person re-identification. InCVPR, 2016. 1, 2 [23] Y. Tian, J. Yan, H. Zhang, Y. Zhang, X. Yang, and H. Zha. Onthe convergence of graph matching: Graduated assignmentrevisited. In ECCV, 2012. 5 [24] T. Wang, S. Gong, X. Zhu, and S. Wang. Person re-identification by video ranking. In ECCV, 2014. 5, 7 [25] Z. Wang, R. Hu, C. Liang, and et al. Zero-shot person re-identification via cross-view consistency. In IEEE TMM,2016. 3 [26] Z. Wang, R. Hu, Y. Yu, and et al. Statistical inference ofgaussian-laplace distribution for person verification. In ACMMM, 2017. 5 [27] Y. Xu, L. Lin, W.-S. Zheng, and X. Liu. Human re-identification by matching compositional template with clus-ter sampling. In ICCV, 2013. 2 [28] J. Yan, M. Cho, H. Zha, X. Yang, and S. M. Chu. Multi-graph matching via affinity optimization with graduated con-sistency regularization. In IEEE TPAMI, 2016. 1, 2 [29] M. Yang, P. Zhu, L. Van Gool, and L. Zhang. Face recogni-tion based on regularized nearest points between image sets.In FG, 2013. 6 [30] Y. Yang, J. Yang, J. Yan, S. Liao, D. Yi, and S. Z. Li. Salientcolor names for person re-identification. In ECCV, 2014. 1 [31] M. Ye, J. Chen, Q. Leng, and et al. Coupled-view basedranking optimization for person re-identification. In MMM,2015. 3 [32] M. Ye, C. Liang, Z. Wang, Q. Leng, and J. Chen. Rankingoptimization for person re-identification via similarity anddissimilarity. In ACM MM, 2015. 3 [33] M. Ye, C. Liang, Y. Yu, and et al. Person re-identificationvia ranking aggregation of similarity pulling and dissimilar-ity pushing. In IEEE TMM, 2016. 1 [34] J. You, A. Wu, X. Li, and W.-S. Zheng. Top-push video-based person re-identification. In CVPR, 2016. 1, 5, 6 [35] Z. Zhang and V. Saligrama. Prism: Person re-identificationvia structured matching. In IEEE TCSVT, 2016. 2 [36] R. Zhao, W. Ouyang, and X. Wang. Unsupervised saliencelearning for person re-identification. In CVPR, 2013. 1, 2, 7,8 [37] L. Zheng, Z. Bie, Y. Sun, J. Wang, C. Su, S. Wang, andQ. Tian. Mars: A video benchmark for large-scale personre-identification. In ECCV, 2016. 1, 5, 6, 7, 8 [38] L. Zheng, Y. Yang, and A. G. Hauptmann. Person re-identification: Past, present and future. ar Xiv, 2016. 1 [39] L. Zheng, Y. Yang, and Q. Tian. SIFT meets CNN: A decadesurvey of instance retrieval. IEEE TPAMI, 2017. 1 [40] X. Zhu, X.-Y. Jing, F. Wu, and H. Feng. Video-based personre-identification by simultaneously learning intra-video andinter-video distance metrics. In IJCAI, 2016. 5, 6