Double-Resolution Attention Network for Person Re-Identification

Jiajie Hu; Chungeng Li; Jubai An; Chao Huang

doi:10.3788/LOP202158.2010019

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 20 >Page 2010019 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 20, 2010019 (2021)

Double-Resolution Attention Network for Person Re-Identification

Jiajie Hu, Chungeng Li^*, Jubai An, and Chao Huang

Author Affiliations

College of Information Science and Technology, Dalian Maritime University, Dalian, Liaoning 116026, China

show less

DOI: 10.3788/LOP202158.2010019 Cite this Article Set citation alerts

Jiajie Hu, Chungeng Li, Jubai An, Chao Huang. Double-Resolution Attention Network for Person Re-Identification[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2010019 Copy Citation Text

show less

Fig. 1. Network structure

Download full size

Fig. 2. Network structure of the channel attention mechanism in this paper

Download full size

Fig. 3. Features extracted from different resolution feature maps. (a) Original input image; (b)(c) feature maps extracted without downsampling; (d)(e) feature maps extracted after downsampling

Download full size

Fig. 4. Channeled attention mechanisms focus on different detailed features. (a) Original input image; (b)--(e) heat maps

Download full size

Fig. 5. Accuracy variation of the algorithm as the epoch increases on different datasets. (a) mAP result analysis; (b) Rank-1 result analysis

Download full size

Model	mAP	Rank-1	Rank-5	Rank-10
Baseline	68.5	85.3	94.0	96.3
DR	68.9	86.4	94.6	96.7
DR+Local	70.6	88.5	95.7	97.3
DR+Local+Att	75.1	91.2	96.3	97.9
DR+Local+Att+BN	82.5	94.4	97.8	98.5
DRANet (3 parts)	83.2	93.9	97.6	98.4
DRANet (6 parts)	79.9	92.7	97.6	98.5
DRANet (4 parts)	82.5	94.4	97.8	98.5

Table 1. Improvement effect of different modules on Market-1501 data set unit: %

Method	mAP	r=1
Bow+kissme^[10]	20.8	44.4
SOMAnet^[19]	47.9	73.9
MSCAN^[20]	57.5	80.8
DML^[21]	68.8	87.7
GLAD^[22]	73.9	89.9
HA-CNN^[23]	75.5	91.2
AlignedReID^[8]	79.3	91.8
PCB+RPP^[5]	81.6	93.8
SCPNet-a^[9]	81.8	94.1
IANet^[24]	83.1	94.4
DRANet (proposed)	82.5	94.4
TriNet^[18] (RK)	81.1	86.7
AACN^[25] (RK)	83.0	88.7
PSE+ECN^[26] (RK)	84.0	90.3
AlignedReID^[8] (RK)	90.7	94.4
DRANet (proposed) (RK)	93.0	95.4

Table 2. Comparison of results on Market-1501 unit: %

Method	mAP	r=1
GAN^[13]	47.1	67.7
SVDNet^[27]	56.8	76.7
MultiScale^[28]	60.6	79.2
HA-CNN^[23]	63.8	80.5
GSRW^[29]	66.4	80.7
SCPNet-a^[9]	68.5	84.8
PCB+RPP^[5]	69.2	83.3
GP-ReID^[30]	72.8	85.2
IANet^[24]	73.4	87.1
DRANet (proposed)	74.1	87.2

Table 3. Comparison of results on DUKEMTMC-REID unit: %

Method	Labeled		Detected
Method	mAP	r=1	mAP	r=1
Pose Transfer^[32]	30.5	33.8	28.2	30.1
PAN^[33]	35.0	36.9	34.0	36.3
SVDNet^[27]	37.8	40.9	37.3	41.5
HA-CNN^[23]	41.0	44.4	38.6	41.7
MGCAM^[34]	50.2	50.1	46.9	46.7
MLFN^[35]	49.2	54.7	47.8	52.8
Mancs^[31]	63.9	69.0	60.5	65.5
DRANet (proposed)	65.3	68.3	61.3	64.1

Table 4. Comparison of results on CUHK03 unit: %

Jiajie Hu, Chungeng Li, Jubai An, Chao Huang. Double-Resolution Attention Network for Person Re-Identification[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2010019

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information