基于多池化融合与背景消除网络的跨数据集行人再识别方法

doi:10.11959/j.issn.1000-436x.2020181

摘要/Abstract

摘要：

现有跨数据集行人再识别方法一般致力于减小2个数据集之间的数据分布差异，忽略了背景信息对识别性能的影响。针对上述问题，提出了一种基于多池化融合与背景消除网络的跨数据集行人再识别方法。为了兼顾全局特征和局部特征，同时实现特征的多细粒度表示，构建了多池化融合网络。为了使监督网络能提取有用的行人前景特征，构建了特征级有监督背景消除网络。采用结合行人分类损失及特征激活损失的多任务学习损失函数，在3个公开行人再识别数据集上对方法进行评估，当MSMT17作为训练集时，Market-1501上的跨数据集识别性能mAP为35.53%，相比ResNet50网络提升了9.24%；DukeMTMC-reID上的跨数据集识别性能mAP为41.45%，相比于ResNet50网络提升了10.72%。与现有方法相比，所提方法具有更优的跨数据集行人再识别性能。

关键词: 行人再识别, 跨数据集, 背景消除, 多池化融合, 深度学习

Abstract:

The existing cross-dataset person re-identification methods were generally aimed at reducing the difference of data distribution between two datasets,which ignored the influence of background information on recognition performance.In order to solve this problem,a cross-dataset person re-ID method based on multi-pool fusion and background elimination network was proposed.To describe both global and local features and implement multiple fine-grained representations,a multi-pool fusion network was constructed.To supervise the network to extract useful foreground features,a feature-level supervised background elimination network was constructed.The final network loss function was defined as a multi-task loss,which combined both person classification loss and feature activation loss.Three person re-ID benchmarks were employed to evaluate the proposed method.Using MSMT17 as the training set,the cross-dataset mAP for Market-1501 was 35.53%,which was 9.24% higher than ResNet50.Using MSMT17 as the training set,the cross-dataset mAP for DukeMTMC-reID was 41.45%,which was 10.72% higher than ResNet50.Compared with existing methods,the proposed method shows better cross-dataset person re-ID performance.

Key words: person re-identification, cross-dataset, background elimination, multi-pool fusion, deep learning

中图分类号:

TP391

李艳凤,张斌,孙嘉,陈后金,朱锦雷. 基于多池化融合与背景消除网络的跨数据集行人再识别方法[J]. 通信学报, 2020, 41(10): 70-79.

Yanfeng LI,Bin ZHANG,Jia SUN,Houjin CHEN,Jinlei ZHU. Cross-dataset person re-identification method based on multi-pool fusion and background elimination network[J]. Journal on Communications, 2020, 41(10): 70-79.

图/表 15

图1

图2

图3

表1

图4

图5

图6

图7

图8

表2

表3

表4

表5

表6

表7

参考文献 27

[1]	CHEN K , CHEN Y , HAN C ,et al. Hard sample mining makes person re-identification more efficient and accurate[J]. Nerocomputing, 2020(382): 259-267.
[2]	SERBETCI A , AKGUL Y S . End-to-end training of CNN ensembles for person re-identification,Pattern Recognition[J]. Pattern Recognition, 2020,104:107319.
[3]	李幼蛟, 卓力, 张菁 ,等. 行人再识别技术综述[J]. 自动化学报, 2018,44(9): 1554-1568.
	LI Y J , ZHUO L , ZHANG J ,et al. Robust resource allocation algorithm for heterogeneous wireless network with SWIPT[J]. ACTA Automatica Sinica, 2018,44(9): 1554-1568.
[4]	GOU M , ZHANG X , RATES-BORRAS A . Person re-identification in appearance impaired scenarios[C]// Proceedings of British Machine Vision Conference. Saarland:DBLP, 2016: 1-14.
[5]	MATSUKAWA T , OKABE T , SUZUKI E . Hierarchical gaussian descriptor for person re-identification[C]// Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2016: 1363-1372.
[6]	KOESTINGER M , HIRZER M , WOHLHART P . Large scale metric learning from equivalence constraints[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2012: 2288-2295.
[7]	ZHENG W S , GONG S , XIANG T . Re-identification by relative distance comparison[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013,35(3): 653-668.
[8]	YI D , LEI Z , LIAO S ,et al. Deep metric learning for person re-identification[C]// International Conference on Pattern Recognition. Stockholm:Institute of Electrical and Electronics Engineers Incorporated, 2014: 34-39.
[9]	YAO H , ZHANG S , ZHANG Y ,et al. Deep representation learning with part loss for person re-identification[J]. IEEE Transactions on Image Processing, 2019,28(6): 2860-2871.
[10]	SUN Y , ZHENG L , YANG Y ,et al. Beyond part models:person retrieval with refined part pooling (and a strong convolutional baseline)[C]// Proceedings of the European Conference on Computer Vision. Berlin:Springer, 2018: 501-508.
[11]	FU Y , WEI Y , ZHOU Y ,et al. Horizontal pyramid matching for person re-identification[J]. arXiv Preprint,arXiv:1804.05275, 2018
[12]	QI L , WANG L , HUO J ,et al. A novel unsupervised camera-aware domain adaptation framework for person re-identification[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway:IEEE Press, 2019: 8079-8088.
[13]	LI Y , LIN C , LIN Y ,et al. Cross-dataset person re-identification via unsupervised pose disentanglement and adaptation[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway:IEEE Press, 2019: 7918-7928.
[14]	HUANG H , YANG W , CHEN X ,et al. EANet:enhancing alignment for cross-domain person re-identification[J]. arXiv Preprint,arXiv:1812.11369, 2018
[15]	WANG J , ZHU X , GONG S ,et al. Transferable joint attribute-identity deep learning for unsupervised person re-identification[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2018: 2275-2284.
[16]	LIN S , LI H , LI C ,et al. Multi-task mid-level feature alignment network for unsupervised cross-dataset person re-identification[J]. arXiv Preprint,arXiv:1807.01440, 2018
[17]	ZHONG Z , ZHENG L , LI S ,et al. Generalizing a person retrieval model hetero-and homogeneously[C]// Proceedings of the European Conference on Computer Vision. Berlin:Springer, 2018: 176-192.
[18]	WEI L , ZHANG S , GAO W ,et al. Person transfer gan to bridge domain gap for person re-identification[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2018: 79-88.
[19]	DENG W , ZHENG L , YE Q ,et al. Image-image domain adaptation with preserved self-similarity and domain-dissimilarity for person re-identification[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2018: 994-1003.
[20]	LIU J , ZHA Z , CHEN D ,et al. Adaptive transfer network for cross-domain person re-identification[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2019: 7202-7211.
[21]	TIAN M , YI S , LI H ,et al. Eliminating background-bias for robust person re-identification[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2018: 5794-5803.
[22]	HE K , ZHANG X , REN S . Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2016: 770-778.
[23]	LONG J , SHELHAMER E , DARRELL T ,et al. Fully convolutional networks for semantic segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2015: 3431-3440.
[24]	LUO Y , ZHENG Z , ZHENG L ,et al. Macro-micro adversarial network for human parsing[C]// Proceedings of the European Conference on Computer Vision. Berlin:Springer, 2018: 424-440.
[25]	GONG K , LIANG X , ZHANG D ,et al. Look into person:self-supervised structure-sensitive learning and a new benchmark for human parsing[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2017: 6757-6765.
[26]	ZHENG L , SHEN L , TIAN L . Scalable person re-identification:a benchmark[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2015: 1116-1124.
[27]	ZHENG Z , ZHENG L , YANG Y . Unlabeled samples generated by GAN improve the person re-identification baseline in vitro[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway:IEEE Press, 2017: 3774-3782.

数据集	网络	Rank-1	Rank-5	Rank-10	mAP
	ResNet50	82.40%	93.15%	95.50%	60.76%
Market-1501	G-L	87.97%	95.63%	97.18%	68.18%
	MPF	91.90%	97.01%	98.25%	76.35%
	ResNet50	74.66%	85.92%	89.79%	54.23%
DukeMTMC-reID	G-L	78.90%	88.78%	91.52%	61.10%
	MPF	83.94%	92.11%	94.21%	68.29%
	ResNet50	60.85%	74.85%	80.68%	30.67%
MSMT17	G-L	66.00%	78.97%	83.38%	34.43%
	MPF	72.07%	83.07%	86.50%	40.70%

训练集	网络	Rank-1	Rank-5	Rank-10	mAP
	ResNet50	47.60%	65.03%	71.80%	20.45%
DukeMTMCreID	MPF	58.11%	73.97%	79.25%	28.04%
	MPF+背景消除	62.48%	78.12%	84.27%	30.72%
	ResNet50	54.93%	71.20%	77.47%	26.29%
MSMT17	MPF	62.33%	78.51%	84.09%	32.22%
	MPF+背景消除	67.85%	82.52%	86.97%	35.53%

训练集	网络	Rank-1	Rank-5	Rank-10	mAP
	ResNet50	32.50%	47.40%	52.52%	16.43%
Market-1501	MPF	43.05%	57.23%	63.51%	24.64%
Market-1501	MPF+背景消除	55.57%	68.81%	74.06%	30.73%
MSMT17	ResNet50	56.06%	70.29%	75.58%	34.20%
MSMT17	MPF	60.09%	74.01	78.54%	36.97%
	MPF+背景消除	64.91%	77.88	82.10%	41.45%

训练集	网络	Rank-1	Rank-5	Rank-10	mAP
	ResNet50	14.62%	23.82%	28.75%	4.31%
Market-1501	MPF	23.92%	34.09%	39.91%	7.87%
Market-1501	MPF+背景消除	32.80%	44.91%	50.45%	10.76%
DukeMTMCreID	ResNet50	22.43%	34.30%	39.88%	6.95%
DukeMTMCreID	MPF	29.35%	41.59%	47.48%	9.36%
	MPF+背景消除	38.65%	51.20%	56.67%	13.19%

数据集	ResNet50/min	MPF/min	MPF+背景消除/min
Market-1501	66	96	200
DukeMTMC-reID	100	150	325
MSMT17	200	240	560