基于风格迁移及度量融合的行人再识别研究
|
摘要
行人再识别是跨场景跨时间的行人图像匹配问题.锁定的目标从一个摄像头下消失后,在其它摄像头视角下再出现时,系统仍能够依据其特征重新锁定.目前该问题遇到的挑战主要来自光照、背景、行人姿态等变化造成的影响.此外,在与训练集不同的数据集上进行测试时性能严重下降,且对大量的数据进行标注的成本非常高.本文采用风格迁移和度量融合的方法:首先,采用循环对抗生成网络将一个数据集中带标签的数据图像风格转换到另一个无标签的数据集上;然后,在风格转换后的数据图像上进行训练,并采用直接度量和间接度量相结合的方式进行相似度度量;最后,在无标签数据集上测试,并将行人图像按相似度由高到低排列输出.实验结果表明:本文方法可明显提高跨数据集的行人再识别准确度.
Person re-identification refers to the problem of pedestrian image matching across multiple scenes and multiple times.When the locked target disappears from a camera and appears at other camera angles,the system can still relock target according to its characteristics.At present,the challenges mainly come from changes in lighting,background,and pedestrian posture.In addition,recognition performance is severely degraded when testing data sets differ from training data sets,and it is very costly to label large amounts of data.The method based on style migration and measurement fusion is adopted in this paper.Firstly,the CycleGAN is utilized to transform the style of data images tagged in one data set to another unlabeled data set.Then,the whole model is trained on style-converted data images and the direct metrics and indirect metrics are combined to measure the similarity.Finally,the testing process is performed on the unlabeled data set and the pedestrian images are output in descending order of similarity.Experimental results show that this method can significantly improve pedestrian re-identification accuracy across data sets.
Person re-identification refers to the problem of pedestrian image matching across multiple scenes and multiple times.When the locked target disappears from a camera and appears at other camera angles,the system can still relock target according to its characteristics.At present,the challenges mainly come from changes in lighting,background,and pedestrian posture.In addition,recognition performance is severely degraded when testing data sets differ from training data sets,and it is very costly to label large amounts of data.The method based on style migration and measurement fusion is adopted in this paper.Firstly,the CycleGAN is utilized to transform the style of data images tagged in one data set to another unlabeled data set.Then,the whole model is trained on style-converted data images and the direct metrics and indirect metrics are combined to measure the similarity.Finally,the testing process is performed on the unlabeled data set and the pedestrian images are output in descending order of similarity.Experimental results show that this method can significantly improve pedestrian re-identification accuracy across data sets.
引文
[1]Gray D,Tao Hai.Viewpoint invariant pedestrian recognition with an ensemble of localized features[C].David Forsyth, Philip Torr, Andrew Zisserman.Computer Vision-ECCV2008,Proceeding of the European Conference on Computer Vision. Marseille France,October 12-18,2008,Springer,Berlin,Heidelberg,2008:262-275.
[2]Zhao Rui,OuYang Wanli,Wang Xiaogang.Unsupervised salience learning for person re-identification[C].Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Portland,USA Portland,June 23-28th,2013:3586-3593.
[3]Prosser B,Zheng Weishi,Gong Shaogang,et al.Person re-identification by support vector ranking[C].Proceedings of the British Machine Vision Conference,Aberystwyth,British,30th August-2nd September,2010:231-242.
[4]Zheng Weishi,Gong Shaogang,Xiang Tao.Person Re-identification by Probabilistic Relative Distance Comparison[C].Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,2011:649-656,IEEE Conference on Computer Vision and Pattern Recognition,CVPR2011.
[5]Isola P,Zhu J Y,Zhou T,et al.Image-to-image translation with conditional adversarial networks[C].CVPR2017,CVPR,Hawaii,July21-26,2017.arXiv preprint arXiv:1611.07004,2016:5967-5976.
[6]Zhu J Y,Park T,Isola P,et al.Unpaired image-toimage translation using cycle-consistent adversarial networks[C].ICCV,Venice,Italy,October 22-29,2017:2242-2251.
[7]Yi Z,Zhang H,Tan P,et al.DualGAN:Unsupervised Dual Learning for Image-to-Image Translation[C].IEEE International Conference on Computer Vision.Venice,Italy,October 22-29,2017.IEEE Computer Society,2017:2868-2876.
[8]Kim T,Cha M,Kim H,et al.Learning to Discover Cross-Domain Relations with Generative Adversarial Networks[C].,2017.
[9]Goodfellow I,Pougetabadie J,Mirza M,et al.Generative adversarial Networks[J].Advances in Neural Information Processing System,2014(3):2672-2680.
[10]Zhu J Y,Park T,Isola P,et al.Unpaired Image-toImage Translation using Cycle-Consistent Adversarial Networks[J].ICCV2017,2017(3):2242-2251.
[11]Li W,Zhao R,Wang X.Human Reidentification with Transferred Metric Learning[C].Proceedings of the Asian Conference on Computer Vision.Berlin,Germany:Springer,2012:31-44.
[12]Xu Y,Zhu Q,Fan Z,et al.Coarse to Fine K Nearest Neighbor Classifier[J].Pattern Recognition Letters,2013,34(9):980-986.
[13]Bai S,Bai X.Sparse contextual activation for efficient visual re-ranking[J].IEEE Transactions on Image Processing,2016,25(3):1056-1069.
[14]Ye M,Liang C,Yu Y,et al.Person reidentification via ranking aggregation of similarity pulling and dissimilarity pushing[J].IEEE Transactions on Multimedia,2016,PP(99):1-1.
[15]Zhong Z,Zheng L,Cao D,et al.Re-ranking Person Reidentification with k-reciprocal Encoding[C].CVPR2017,CVPR,Hawaii,July21-26,2017.IEEE Xplore,2017:3652-3661.
[2]Zhao Rui,OuYang Wanli,Wang Xiaogang.Unsupervised salience learning for person re-identification[C].Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Portland,USA Portland,June 23-28th,2013:3586-3593.
[3]Prosser B,Zheng Weishi,Gong Shaogang,et al.Person re-identification by support vector ranking[C].Proceedings of the British Machine Vision Conference,Aberystwyth,British,30th August-2nd September,2010:231-242.
[4]Zheng Weishi,Gong Shaogang,Xiang Tao.Person Re-identification by Probabilistic Relative Distance Comparison[C].Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,2011:649-656,IEEE Conference on Computer Vision and Pattern Recognition,CVPR2011.
[5]Isola P,Zhu J Y,Zhou T,et al.Image-to-image translation with conditional adversarial networks[C].CVPR2017,CVPR,Hawaii,July21-26,2017.arXiv preprint arXiv:1611.07004,2016:5967-5976.
[6]Zhu J Y,Park T,Isola P,et al.Unpaired image-toimage translation using cycle-consistent adversarial networks[C].ICCV,Venice,Italy,October 22-29,2017:2242-2251.
[7]Yi Z,Zhang H,Tan P,et al.DualGAN:Unsupervised Dual Learning for Image-to-Image Translation[C].IEEE International Conference on Computer Vision.Venice,Italy,October 22-29,2017.IEEE Computer Society,2017:2868-2876.
[8]Kim T,Cha M,Kim H,et al.Learning to Discover Cross-Domain Relations with Generative Adversarial Networks[C].,2017.
[9]Goodfellow I,Pougetabadie J,Mirza M,et al.Generative adversarial Networks[J].Advances in Neural Information Processing System,2014(3):2672-2680.
[10]Zhu J Y,Park T,Isola P,et al.Unpaired Image-toImage Translation using Cycle-Consistent Adversarial Networks[J].ICCV2017,2017(3):2242-2251.
[11]Li W,Zhao R,Wang X.Human Reidentification with Transferred Metric Learning[C].Proceedings of the Asian Conference on Computer Vision.Berlin,Germany:Springer,2012:31-44.
[12]Xu Y,Zhu Q,Fan Z,et al.Coarse to Fine K Nearest Neighbor Classifier[J].Pattern Recognition Letters,2013,34(9):980-986.
[13]Bai S,Bai X.Sparse contextual activation for efficient visual re-ranking[J].IEEE Transactions on Image Processing,2016,25(3):1056-1069.
[14]Ye M,Liang C,Yu Y,et al.Person reidentification via ranking aggregation of similarity pulling and dissimilarity pushing[J].IEEE Transactions on Multimedia,2016,PP(99):1-1.
[15]Zhong Z,Zheng L,Cao D,et al.Re-ranking Person Reidentification with k-reciprocal Encoding[C].CVPR2017,CVPR,Hawaii,July21-26,2017.IEEE Xplore,2017:3652-3661.