计算机系统应用  2020, Vol. 29 Issue (1): 256-260 PDF

Semi-Supervised Person Re-Identification with One-Sample
SHAN Chun, WANG Min
College of Computer and Information, HOHAI University, Nanjing 211100, China
Abstract: In this study, we propose a one-sample person re-identification method, which adopts a progressive learning framework in the process of iteration in order to making full use of the characteristics of labeled data and unlabeled data to optimize the model. In this framework, we iteratively train convolutional neural network to update the model and utilize multiple-model training together to select the reliable pseudo-label data during label estimation. Then, we update training data for the next round of training. The training data is splited into three parts: labeled data, pseudo-labeled data, and indexed-labeled data. We set up the corresponding loss function for each set of data and update the CNN model by the joint training on the three parts. In the progress of iteration, the pseudo-labeled data and index-labeled data are constantly updated. Under the one-sample set, rank-1=65.3, mAP=45.6. When the rate of labeled data is increased to 40%, rank-1=83.8, mAP=64.9. The result indicates that the semi-supervised person re-identification method proposed in this study can provide excellent results comparable to the supervised learning method with less labeled data, which fully demonstrates the effectiveness of the method.
Key words: semi-supervised     one-sample     progressive-learning     multi-model     joint-learning

1 引言

2 半监督式单样本深度行人重识别

 图 1 总体框架

2.1 联合学习方法

 $U = \left\{ {{x_{{n_l} + 1}},{x_{{n_l} + 2}}, \cdots ,{x_{{n_l} + u}}} \right\} = {P^k} + {I^k}$

 ${l_e}\left( {{\rm{V}};\tilde \phi \left( {{\bf{\theta }};{x_i}} \right)} \right) = - {\log}\dfrac{{{\exp}\left( {\frac{{v_i^{\rm T}\tilde \phi \left( {{\bf{\theta }};{x_i}} \right)}}{\tau }} \right)}}{{\displaystyle \sum \nolimits_{j = 1}^{\left| {{M^t}} \right|} {\exp}\left( {\dfrac{{v_i^{\rm T}\tilde \phi \left( {{\bf{\theta }};{x_i}} \right)}}{\tau }} \right)}}$

 ${{{v}}_i} \leftarrow \frac{1}{2}({{{v}}_i} + \tilde \phi \left( {{\bf{\theta }};{x_i}} \right)$

 $\mathop {\min }\limits_{{\bf{\theta }},\omega } \mathop \sum \limits_{i = 1}^{{n_l}} {l_{CE}}\left( {f\left( {\omega ;\phi \left( {{\bf{\theta }};{x_i}} \right)} \right),{y_i}} \right)$

 $\mathop {\min }\limits_{{{\theta}} ,\omega } \mathop \sum \limits_{i = {n_l} + 1}^{{n_l} + {n_u}} {s_i}{l_{CE}}\left( {f\left( {\omega ;\phi \left( {{{\theta}} ;{x_i}} \right)} \right),{y_i}} \right)$

 \begin{aligned} &\mathop {\min }\limits_{{{\theta}} ,\omega } c\mathop \sum \limits_{i = 1}^{{n_l}} {s_i}{l_{CE}}\left( {f\left( {\omega ;\phi \left( {{{\theta}} ;{x_i}} \right)} \right),{y_i}} \right)\\ & + c\mathop \sum \limits_{i = {n_l} + 1}^{{n_l} + {n_u}} s_i^{k - 1}{l_{CE}}\left( {f\left( {\omega ;\phi \left( {{{\theta}} ;{x_i}} \right)} \right),{y_i}} \right)\\ & + \left( {1 - c} \right)\mathop \sum \limits_{i = {n_l} + 1}^{{n_l} + {n_u}} \left( {1 - s_i^{k - 1}} \right){l_e}(f\left( {V;\tilde \phi \left( {{{\theta}} ;{x_i}} \right)} \right) \end{aligned}

k表示第k次迭代, c为调整交叉熵损失与差异性损失所占权重的参数.

2.2 抽样策略与标签估计

 \left\{\begin{aligned} & d\left( {{\bf{\theta }};{x_i}} \right) = \left| {\left| {\phi \left( {{\bf{\theta }};{x_i}} \right) - \phi \left( {{\bf{\theta }};x'} \right)} \right|} \right|\\ &x',y' = {\rm{arg}}\mathop {\min }\limits_{\left( {{x_l},{y_l}} \right) \in L} \left| {\left| {\phi \left( {{\bf{\theta }};{x_i}} \right) - \phi \left( {{\bf{\theta }};x_l'} \right)} \right|} \right|\\ &{y_i} = y' \end{aligned}\right.

$d\left( {\varphi ;{x_i}} \right)$ 看作是提取特征向量的不相似代价, 用于估计样本标签. 通过上一步, 可以从两个模型中各自挑选出Nk个置信度较高的伪标签数据. 但是, 这样的伪标签的数据标签精确度仍然有限, 为此, 将两个模型挑选出的2Nk个伪标签数据共同进行进一步的筛选, 将特征索引相同的伪标签数据挑选出来, 从而完成了为标签数据精度的再一次提升:

 ${P^t} = P_{{\phi _1}}^t \cap P_{{\phi _2}}^t$
2.3 总体迭代策略

(1) 初始化两个不同的CNN模型, 抽调伪标签数据 $\scriptstyle{P_0} \leftarrow 0$ , 设置抽调样本扩充因数α;

(2) 将两个不同的CNN模型同时进行训练, 在迭代进行过程中, 通过3组数据优化模型参数;

(3) 对无标签数据进行伪标签评估;

(4) 通过两个模型, 选取较优伪标签数据;

(5) 更新抽样规模;

(6) 评估模型阶段, 通过在验证集上的表现, 挑选出两个不同的最优模型;

(7) 更新下次迭代使用的伪标签数据和映射标签数据.

3 实验与分析 3.1 实验数据与评价标准

3.2 实验环境及参数配置

3.3 实验比较

4 结论

 [1] Zheng WS, Gong SG, Xiang T. Person re-identification by probabilistic relative distance comparison. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Providence, RI, USA. 2011. 649–656. [2] Chen YC, Zheng WS, Lai JH, et al. An asymmetric distance model for cross-view feature mapping in person reidentification. IEEE Transactions on Circuits and Systems for Video Technology, 2017, 27(8): 1661-1675. DOI:10.1109/TCSVT.2016.2515309 [3] Gong S, Cristani M, Yan S. Person re-identification. Computer Vision and Pattern Recognition, 2014, 42(7): 301-303. [4] Li DW, Chen XT, Zhang Z, et al. Learning deep context-aware features over body and latent parts for person re-identification. The IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA. 2017. 384–393. [5] Wojke N, Bewley A. Deep cosine metric learning for person re-identification. IEEE Winter Conference on Applications of Computer Vision. Lake Tahoe, NV, USA. 2018. 748–756. [6] Huang Y, Xu JS, Wu Q, et al. Multi-pseudo regularized label for generated data in person re-identification. IEEE Transactions on Image Processing, 2019, 28(3): 1391-1403. DOI:10.1109/TIP.2018.2874715 [7] Kipf TN, Welling M. Semi-supervised classification with graph convolutional networks. arXiv: 1609.02907, 2017. [8] Liu X, Song ML, Tao DC, et al. Semi-supervised coupled dictionary learning for person re-identification. IEEE Conference on Computer Vision and Pattern Recognition, 2014, 3550-3557. [9] Ding GD, Zhang SS, Khan S, et al. Center based pseudo-labeling for semi-supervised person re-identification. IEEE International Conference on Multimedia & Expo Workshops. San Diego, CA, USA. 2018. 1–6. [10] Li JW, Ma AJ, Yuen PC. Semi-supervised region metric learning for person re-identification. International Journal of Computer Vision, 2018, 126(8): 855-874. DOI:10.1007/s11263-018-1075-5 [11] Chen F, Chai JH, Ren DH, et al. Semi-supervised distance metric learning for person re-identification. 2017 IEEE/ACIS 16th International Conference on Computer and Information Science. Wuhan, China. 2017. 733–738. [12] Wu Y, Li YT, Dong XY, et al. Exploit the unknown gradually: One-shot video-based person re-identification by stepwise learning. IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA. 2018. 5177–5186. [13] Zheng L, Bie Z, Sun YF, et al. MARS: A video benchmark for large-scale person re-identification. European Conference on Computer Vision. Cham, 2016, 868-884. [14] Szegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the inception architecture for computer vision. IEEE Conference on Computer Vision and Pattern Recognition, 2016, 2818-2826. [15] He KM, Zhang XY, Ren SQ, et al. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, USA. 2016. 770–778. [16] Ye M, Ma AJ, Zheng L, et al. Dynamic label graph matching for unsupervised video re-identification. 2017 IEEE International Conference on Computer Vision. Venice, Italy. 2017. 5142–5150.