Domain-Adaptive Person Search With Diverse Images and Instance Augmentation

Fig. 4. Visualization of changing trends during training. (a) mAP; (b) number of pseudo persons in memory bank; (c) number of target-domain false positives; (d) number of source-domain negatives

Download full size

Fig. 5. Examples of diverse negatives mined in target domain

Download full size

Fig. 6. Examples of diverse negatives mined in source domain

Download full size

Table 1. Comparison with state-of-the-art methods on CUHK-SYSU and PRW test sets

View table

Table 1. Comparison with state-of-the-art methods on CUHK-SYSU and PRW test sets

Method		CUHK-SYSU		PRW
Method		mAP /%	Top-1 /%	mAP /%	Top-1 /%
	NPSM^［22］	77.9	81.2	24.2	53.1
	CLSA^［8］	87.2	88.5	38.7	65.0
	RCAA^［23］	79.3	81.3
	IAN^［24］	76.3	80.1	23.0	61.9
	CTXG^［25］	84.1	86.5	33.4	73.6
	QEEPS^［26］	88.9	89.1	37.1	76.7
Traditional fully-supervised approach	HOIM^［27］	89.7	90.8	39.8	80.4
	RDLR^［28］	93.0	94.2	42.9	70.2
	APNet^［29］	88.9	89.3	41.9	81.4
	BINet^［30］	90.0	90.7	45.3	81.7
	NAE^［10］	91.5	92.4	43.3	80.9
	NAE+^［10］	92.1	92.9	44.0	81.1
	PGSFL^［31］	90.2	91.8	42.5	83.5
	SeqNet^［19］	93.8	94.6	46.7	83.4
	AlignPS^［15］	93.1	93.4	45.9	81.9
	SeqNet^［19］	52.5	54.8	30.3	77.7
	AlignPS^［15］	39.7	44.0	27.9	77.4
Domain-adaptive approach	DAPS^［18］	77.6	79.6	34.7	80.6
	DIIA （ours）	80.0	81.4	40.8	81.0

Table 2. Comparison of cross domain re-identification accuracy and speed of person search methods on the CUHK-SYSU dataset
View table
Table 2. Comparison of cross domain re-identification accuracy and speed of person search methods on the CUHK-SYSU dataset
Method mAP /% Time /ms
SeqNet^［19］ 52.5 86
AlignPS^［15］ 39.7 61
DAPS^［18］ 77.6 82
DIIA （ours） 80.0 82

Table 3. Influence of different components when source domain is PRW and target domain is CUHK-SYSU
View table
Table 3. Influence of different components when source domain is PRW and target domain is CUHK-SYSU
Number SIA NRL mAP /% Top-1 /% AP /%
① 77.6 79.6 68.3
② √ 78.3 80.1 69.7
③ √ √ 80.0 81.4 70.1

Table 4. Influence of different components when source domain is CUHK-SYSU and target domain is PRW
View table
Table 4. Influence of different components when source domain is CUHK-SYSU and target domain is PRW
Number SIA NRL mAP/% Top-1/% AP/%
① 34.7 80.6 89.9
② √ 37.3 80.6 90.2
③ √ √ 40.8 81.0 90.3

Table 5. Influence of image augmentation on different domains
View table
Table 5. Influence of image augmentation on different domains
Number Method mAP /% Top-1 /%
① No image augmentation 77.6 79.6
② Target-domain image augmentation 69.7 71.4
③ Dual-domain image augmentation 71.5 72.9
④ Source-domain image augmentation 78.3 80.1

Table 6. Influence of different image augmentation strategies on our SIA module
View table
Table 6. Influence of different image augmentation strategies on our SIA module
Number Method mAP /% Top-1 /%
① No image augmentation 77.6 79.6
② Only Gaussian blur 77.9 79.8
③ Only color jitter 78.0 79.6
④ Only elastic transform 78.1 79.9
⑤ All 78.3 80.1

Table 7. Influence of two strategies in DNM module
View table
Table 7. Influence of two strategies in DNM module
Number Method mAP /% Top-1 /%
① SIA 78.3 80.1
② SIA+TFPR 79.3 80.6
③ SIA+SNDU 79.0 80.3
④ SIA+TFPR+SNDU 80.0 81.4

Table 8. Influence of number of recycled false positives in TFPR
View table
Table 8. Influence of number of recycled false positives in TFPR
Number Method mAP /% Top-1 /%
① Without TFPR 78.3 80.1
② Last 10 false positives 78.9 80.5
③ Same number of positives 79.3 80.6

Table 9. Influence of threshold setting in SNDU
View table
Table 9. Influence of threshold setting in SNDU
Number Method mAP /% Top-1 /%
① Without SNDU 78.3 80.1
② Fixed threshold 0.8 78.6 80.3
③ Dynamic threshold 79.0 80.3

Tools

Get Citation

Copy Citation Text

Zhiqiang Dong, Jiale Cao, Aiping Yang. Domain-Adaptive Person Search With Diverse Images and Instance Augmentation[J]. Laser & Optoelectronics Progress, 2024, 61(14): 1437011

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Nov. 30, 2023

Accepted: Jan. 8, 2024

Published Online: Jul. 8, 2024

The Author Email: Jiale Cao (connor@tju.edu.cn)

DOI:10.3788/LOP232590

CSTR:32186.14.LOP232590

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparison with state-of-the-art methods on CUHK-SYSU and PRW test sets

Table 1. Comparison with state-of-the-art methods on CUHK-SYSU and PRW test sets

Table 2. Comparison of cross domain re-identification accuracy and speed of person search methods on the CUHK-SYSU dataset

Table 2. Comparison of cross domain re-identification accuracy and speed of person search methods on the CUHK-SYSU dataset

Table 3. Influence of different components when source domain is PRW and target domain is CUHK-SYSU

Table 3. Influence of different components when source domain is PRW and target domain is CUHK-SYSU

Table 4. Influence of different components when source domain is CUHK-SYSU and target domain is PRW

Table 4. Influence of different components when source domain is CUHK-SYSU and target domain is PRW

Table 5. Influence of image augmentation on different domains

Table 5. Influence of image augmentation on different domains

Table 6. Influence of different image augmentation strategies on our SIA module

Table 6. Influence of different image augmentation strategies on our SIA module

Table 7. Influence of two strategies in DNM module

Table 7. Influence of two strategies in DNM module

Table 8. Influence of number of recycled false positives in TFPR

Table 8. Influence of number of recycled false positives in TFPR

Table 9. Influence of threshold setting in SNDU

Table 9. Influence of threshold setting in SNDU