基于奇异值分解行人对齐网络的行人重识别

doi:10.6040/j.issn.1672-3961.0.2018.347

摘要/Abstract

摘要：

为解决行人重识别的训练数据集中自动检测出的行人图像背景过大和行人部分缺失的错位现象问题,使用空间变换网络层对图像错位进行处理。为优化整个网络的深度学习过程,提高图像检索能力,增加网络特征层,使用奇异值向量分解方法对其进行处理。将行人对齐网络和奇异向量分解相结合,构造奇异值分解行人对齐网络,既可解决图像错位问题,又提高图像特征的相似性度量的效果。在Market1501、CUHK03和DukeMTMC-reID数据集上进行试验,并与行人对齐网络和其他深度学习与非深度学习的行人重识别方法进行比较,试验结果中整个网络的平均检索精度和行人图像第一次匹配正确的概率平均达到了65%和80%左右,这表明奇异值分解行人对齐网络可以提高对行人匹配的效果。

关键词: 行人重识别, 奇异向量分解, 行人对齐, 深度学习, 空间变换网络

Abstract:

In order to solve the problem that the background of pedestrian image was too large and the part of pedestrian was missing in the training data set of pedestrian recognition, the spatial transformation network layer was used to process the image dislocation. In order to optimize the deep learning process of the whole network and improve the image retrieval ability, a feature layer was added to the network, and singular value vector decomposition was used to process it. By combining the pedestrian alignment network with the singular vector decomposition and constructing the singular value decomposition pedestrian alignment network, the image dislocation problem could be solved and the effect of similarity measurement of image features could be improved. Experiments were conducted on Market1501, CUHK03 and DukeMTMC-reID datasets, and compared with pedestrian alignment network and other pedestrian re-recognition methods of deep learning and non-deep learning. In the experimental results, the values of rank-1 and mAP mean average precision reached 80% and 65% on average, which indicated that singular value decomposition of pedestrian alignment network had certain benefits on pedestrian matching effect.

Key words: pedestrian recognition, singular vector decomposition, pedestrian alignment, deep learning, spatial transformation network

中图分类号:

TP391

张继,金翠,王洪元,陈首兵. 基于奇异值分解行人对齐网络的行人重识别[J]. 山东大学学报 (工学版), 2019, 49(5): 91-97.

Ji ZHANG,Cui JIN,Hongyuan WANG,Shoubing CHEN. Pedestrian recognition based on singular value decomposition pedestrian alignment network[J]. Journal of Shandong University(Engineering Science), 2019, 49(5): 91-97.

图/表 6

图1

表1

表2

表3

表4

表5

参考文献 21

1	ZHENG Liang, SHEN Liyue, TIAN Lu, et al. Scalable person re-identification: a benchmark[C]//IEEE International Conference on Computer Vision. Santiago, Chile: IEEE Computer Society, 2015: 1116-1124.
2	WANG Hongyuan , DING Zongyuan , ZHANG Ji , et al. Person reidentification by semisupervised dictionary rectification learning with retraining module[J]. Journal of Electronic Imaging, 2018, 27 (4): 043043-1- 9.
3	NI Tongguang , DING Zongyuan , CHEN Fuhua , et al. Relative distance metric leaning based on clustering centralization and projection vectors learning for person re-identification[J]. IEEE Access, 2018, 6 (1): 11405- 11411.
4	丁宗元, 王洪元, 陈付华, 等. 基于距离中心化与投影向量学习的行人重识别[J]. 计算机研究与发展, 2017, 54 (8): 1785- 1794.
	DING Zongyuan , WANG Hongyuan , CHEN Fuhua , et al. Pedestrian weight recognition based on distance centralization and projection vector learning[J]. Computer Research and Development, 2017, 54 (8): 1785- 1794.
5	WANG Hongyuan, DING Zongyuan, NI Tongguang, et al. KL divergence based person re-identification using multivariate Gaussian distributions[C]//Proceedings of the 2017 4th Asian Conference on Pattern Recognition. Nanjing, China: Sponsorship, 2017: 417-422.
6	张文文, 王洪元, 万建武, 等. 基于稀疏学习的行人重识别算法[J]. 数据采集与处理, 2018, 33 (5): 855- 864.
	ZHANG Wenwen , WANG Hongyuan , WAN Jianwu , et al. Pedestrian weight recognition algorithm based on sparse learning[J]. Data Acquisition and Processing, 2018, 33 (5): 855- 864.
7	NI Tongguang , GU Xiaoqing , WANG Honyuan , et al. Discirminative deep trasfer metric learning for cross-scenario person re-identification[J]. Journal of Electronic Imaging, 2018, 27 (4): 043026- 1.
8	陈首兵, 王洪元, 金翠, 等. 基于孪生网络和重排序的行人重识别[J]. 计算机应用, 2018, 38 (11): 3161- 3166. doi: 10.11772/j.issn.1001-9081.2018041223
	CHEN Shoubing , WANG Hongyuan , JIN Cui , et al. Pedestrian weight recognition based on twin networks and reordering[J]. Computer Applications, 2018, 38 (11): 3161- 3166. doi: 10.11772/j.issn.1001-9081.2018041223
9	金翠, 王洪元, 陈首兵. 基于随机擦除行人对齐网络的行人重识别方法[J]. 山东大学学报(工学版), 2018, 48 (6): 67- 73.
	JIN Cui , WANG Hongyuan , CHEN Shoubing . Pedestrian weight recognition method based on random erasure pedestrian alignment network[J]. Journal of Shandong University(Engineering Science Edition), 2018, 48 (6): 67- 73.
10	DENTON E, ZAREMBA W, BRUNA J, et al. Exploiting linear structure within convolutional networks for efficient evaluation[C]//Proceedings of the Conference and Workshop on Neural Information Processing Systems. Montréal, Canada: IEEE Computer Society, 2014: 1269-1277.
11	ZHAO Liming, LI Xi, WANG Jingdong, et al. Deeply-learned part-aligned representations for person re-identification[C]//Proceedings of the International Conference on Computer Vision. Venice, Italy: IEEE Computer Society, 2017: 3219-3228.
12	ZHENG Zhedong, ZHENG Liang, YANG Yi. Pedestrian alignment network for large-scale person re-identification[J/OL].[2018-12-20]. https://arxiv.org/pdf/1707.00408.pdf.
13	SUN Yifan, ZHENG Liang, DENG Weijian. SVDnet for pedestrian retrieval[C]//Proceedings of the International Conference on Computer Vision. Venice, Italy: IEEE Computer Society, 2017: 3820-3828.
14	SU Chi, ZHANG Shiliang, XING Junliang, et al. Deep attributes driven multi-camera person re-identification[C]//Proceedings of the European Conference on Computer Vision. Amsterdam, the Netherlands: IEEE Computer Society, 2016: 475-491.
15	SARFRAZ M Saquib, SCHUMANN Arne, EBERLE Andreas, et al. A pose-sensitive embedding for person re-identification with expanded cross neighborhood re-ranking[C]//Proceedings of the IEEE Computer Vision and Pattern Recognition(CVPR). Salt Lake City, Utah: IEEE Computer Society, 2018: 420-429.
16	GUO Yiluan, CHEUNG Ngai-man.. Efficient and Deep person re-identification using multi-level similarity[C]//Proceedings of the IEEE Computer Vision and Pattern Recognition(CVPR). Salt Lake City, USA: IEEE Computer Society, 2018: 2335-2344.
17	HE Lingxiao, LIANG Jian, LI Haiqing, et al. Deep spatial feature reconstruction for partial person re-identification: alignment-free approach[C]//Proceedings of the IEEE Computer Vision and Pattern Recognition(CVPR). Salt Lake City, Utah: IEEE Computer Society, 2018: 7073-7082.
18	ZHONG Zhun, ZHENG Liang, CAO Donglin, et al. Re-ranking person re-identification with k-reciprocal encoding[C]//Proceedings of the IEEE Computer Vision and Pattern Recognition(CVPR). Honolulu, Hawaii: IEEE Computer Society, 2017: 3652-3661.
19	LIAO Shengcai, HU Yang, ZHU Xiangyu, et al. Person re-identification by local maximal occurrence representation and metric learning[C]//Proceedings of the IEEE Computer Vision and Pattern Recognition(CVPR). Boston, Massachusetts: IEEE Computer Society, 2015: 2197-2206.
20	VARIOR Rahulrama, SHUAI Bing, LU Jiwen, et al. A siamese long short-term memory architecture for human reidentification[C]//Proceedings of the European Conference on Computer Vision, . Amsterdam, the Netherlands: IEEE Computer Society, 2016: 135-153.
21	VARIOR Rahulrama, HALOI Mrinal, WANG Gang. Gated siamese convolutional neural network architecture for human reidentification[C]//Proceedings of the European Conference on Computer Vision. Amsterdam, the Netherlands: IEEE Computer Society, 2016: 791-808.

多维度评价

Viewed

Full text

Abstract

Cited

Shared

Discussed

网络	方法	Market1501数据集		DukeMTMC-reID数据集		CUHK03(labeled)数据集		CUHK03(detected)数据集
网络	方法	rank-1/%	mAP/%	rank-1/%	mAP/%	rank-1/%	mAP/%	rank-1/%	mAP/%
第1条分支	PAN	80.17	59.14	65.2	44.99	31.14	29.80	30.50	29.04
第1条分支	ours	81.38	62.02	66.74	45.96	79.19	78.82	77.06	80.29
第3条分支	PAN	79.01	58.27	69.36	47.14	35.29	32.90	34.14	31.71
第3条分支	ours	80.17	60.85	69.07	48.86	78.55	79.12	78.68	81.41
整个网络	PAN	82.81	63.35	68.36	51.51	36.86	35.03	36.29	34.05
整个网络	ours	83.41	65.09	75.54	52.20	79.63	80.45	78.35	81.22

%
方法	Market1501	DuckMTMC-reID	CUHK03(labeled)	CUHK03(detected)
SVDNet	82.3	80.05	81.80	76.70
ours	83.41	75.54	79.63	78.35

%
方法	Market1501	DuckMTMC-reID	CUHK03(labeled)	CUHK03(detected)
PAN	82.81	80.05	36.86	36.29
ours	83.41	75.54	79.63	78.35

%
方法	Market1501		CUHK03
方法	rank-1	mAP	rank-1	mAP
LOMO+XQDA	43.80	22.20	44.60	51.50
Siamese LSTM	61.60	35.30	57.30	46.30
Gated SCNN	65.90	39.60	61.80	51.30
DLCE	79.50	59.90	79.63	80.45
ours	83.41	65.09	79.52	81.34

%
方法	rank-1	mAP
Bow+kissme	25.1	12.2
LOMO+XQDA	30.8	17
Baseline	65.5	44.1
GAN	67.7	47.1
ours	75.54	52.2