Person Re-Identification Based on Pose Estimation with Feature Similarity

Rui Li; Min Jiang

doi:10.3788/LOP212869

Laser & Optoelectronics Progress, Volume. 60, Issue 6, 0610001(2023)

Person Re-Identification Based on Pose Estimation with Feature Similarity

Rui Li and Min Jiang^*

Engineering Laboratory of Pattern Recognition and Computational Intelligence, School of Artificial Intelligence and Computer, Jiangnan University, Wuxi 214122, Jiangsu, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Structure of PAMFNet

Download full size

Fig. 2. Results of PRAP. (a) Original pedestrian images; (b) pedestrian key points extracted by pose estimation algorithm; (c) local feature pooling region based on attitude estimation; (d) aligned upper and lower body global features

Download full size

Fig. 3. Structure of similarity matrix

Download full size

Fig. 4. Thermodynamic diagrams of feature. (a) Original pedestrian image; (b) responsed thermodynamic diagram

Download full size

Fig. 5. Top10 search results of partial query images

Download full size

Table 1. Comparison to state-of-the-art methods on Market-1501 dataset

View table

Table 1. Comparison to state-of-the-art methods on Market-1501 dataset

Method	Market1501
Method	Rank-1 /%	mAP /%
PCB+RPP［4］	93.8	81.6
AANet［19］	93.9	83.4
Pyramid［20］	95.7	88.2
MGN［21］	95.7	86.9
DSAP［22］	95.7	87.6
CAMA［23］	94.7	84.5
OSNet［24］	94.8	84.9
MHN［25］	95.1	85.0
FPR［26］	95.4	86.6
Ours	95.4	88.7

Table 2. Comparison to state-of-the-art methods on DukeMTMC-ReID dataset

View table

Table 2. Comparison to state-of-the-art methods on DukeMTMC-ReID dataset

Method	DukeMTMC-ReID
Method	Rank-1 /%	mAP /%
PCB+RPP［4］	83.3	69.2
AANet［19］	87.7	74.3
Pyramid［20］	89.0	79.0
MGN［21］	88.7	78.4
DSAP［22］	86.2	74.3
CAMA［23］	85.8	72.9
OSNet［24］	88.6	73.5
MHN［25］	89.1	77.2
FPR［26］	88.6	78.4
Ours	88.7	78.9

Table 3. Comparison to state-of-the-art methods on CUHK03 dataset

View table

Table 3. Comparison to state-of-the-art methods on CUHK03 dataset

Method	CUHK03
	Labeled		Detected
	Rank-1 /%	mAP /%	Rank-1 /%	mAP /%
PCB+RPP［4］	-	-	63.7	57.5
MGN［21］	68.0	67.4	68.0	66.0
CAMA［23］	70.1	66.5	66.6	64.2
Auto-ReID［27］	77.9	73.0	73.3	69.3
MHN［25］	77.2	72.4	71.7	65.4
FPR［26］	-	-	76.1	72.3
Ours	78.5	75.6	76.2	73.7

Table 4. Comparison to state-of-the-art methods on MSMT17 dataset
View table
Table 4. Comparison to state-of-the-art methods on MSMT17 dataset
Method MSMT17
Rank-1 /% mAP /%
IANet［28］ 75.5 45.8
DGNet［29］ 77.2 52.3
Auto-ReID［27］ 78.2 52.5
OSNet［24］ 78.7 52.9
Ours 80.1 56.3

Table 5. Effect of PRAP
View table
Table 5. Effect of PRAP
Method Market1501 Duke
Rank-1 /% mAP /% Rank-1 /% mAP /%
Baseline 92.3 77.4 81.7 66.1
Baseline+PRAP_5P 93.9 86.5 84.3 72.4
Baseline+PRAP_8P（Ours） 94.3 86.8 85.0 74.8
Baseline+PRAP_10P 94.1 86.2 84.7 72.9

Table 6. Effect of part feature branch
View table
Table 6. Effect of part feature branch
Method Market1501 Duke
Rank-1 /% mAP /% Rank-1 /% mAP /%
Baseline 92.3 77.4 81.7 66.1
Baseline+PRAP+s3 94.4 87.8 86.9 77.8
Baseline+PRAP+s4（Ours） 94.9 87.9 87.5 78.7
Baseline+PRAP+s3+s4 94.7 87.3 87.6 78.5

Table 7. Effect of SSM
View table
Table 7. Effect of SSM
Method Market1501 Duke
Rank-1 /% mAP /% Rank-1 /% mAP /%
Baseline 92.3 77.4 81.7 66.1
Baseline+PRAP+fore 95.1 88.3 87.9 79.0
Baseline+PRAP+fore+triLoss 95.4 88.7 88.7 78.9

Tools

Get Citation

Copy Citation Text

Rui Li, Min Jiang. Person Re-Identification Based on Pose Estimation with Feature Similarity[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0610001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites