Cross-Modal Pedestrian Re-Identification Combining Frequency-Domain Attention and Modal Co-Feature Optimization

Method	Rank-1 accuracy		Rank-10 accuracy		mAP
Method	All-search	Indoor-search	All-search	Indoor-search	All-search	Indoor-search
Zeropadding^［11］	14.80	20.58	54.12	68.38	15.95	26.92
JSIA-ReID^［8］	38.10	43.80	80.70	86.20	36.90	52.90
AlignGAN^［20］	42.40	45.90	85.00	87.60	40.70	54.30
AGW^［14］	47.50	54.17	84.39	91.14	47.65	62.97
DDAG^［17］	54.75	61.02	90.39	94.06	53.02	67.98
FMCNet^［4］	66.34	73.44	—	—	62.51	56.06
MDRA^［19］	70.55	72.69	94.90	97.15	63.89	77.14
MPANet^［30］	70.58	76.74	96.10	98.21	68.24	80.95
MMN^［31］	70.60	76.20	96.20	97.20	66.90	79.60
MCJA^［32］	74.48	82.79	96.99	98.88	71.34	85.26
DEEN^［13］	74.70	80.30	97.60	99.00	71.80	83.30
MUN^［29］	76.24	79.42	97.84	98.09	73.81	82.06
IDKL^［26］	81.42	87.14	97.38	98.28	79.85	89.37
Proposed	82.14	89.35	97.98	98.79	81.59	90.85

Table 2. Performances comparison on the RegDB dataset by different methods

View table

Table 2. Performances comparison on the RegDB dataset by different methods

Method	Rank-1 accuracy		Rank-10 accuracy		mAP
Method	Visible to infrared	Infrared to visible	Visible to infrared	Infrared to visible	Visible to infrared	Infrared to visible
Zeropadding^［11］	17.80	16.60	34.20	34.70	18.90	17.80
JSIA-ReID^［8］	48.50	48.10	—	—	49.30	48.90
AlignGAN^［20］	57.90	56.30	—	—	53.60	53.40
AGW^［14］	70.05	70.49	86.21	87.21	67.64	65.90
DDAG^［17］	69.34	68.06	86.19	85.15	63.46	61.80
FMCNet^［4］	89.10	88.40	—	—	84.40	83.90
MDRA^［19］	91.80	—	97.46	—	82.08	—
MPANet^［30］	83.70	82.80	—	—	80.90	80.70
MMN^［31］	91.60	87.50	97.70	96.00	84.10	80.50
MCJA^［32］	91.80	88.06	86.08	—	—	83.06
DEEN^［13］	91.10	89.50	97.80	96.80	85.10	83.40
MUN^［29］	95.19	91.86	—	—	87.15	85.01
IDKL^［26］	94.72	94.22	—	—	90.19	90.43
Proposed	95.39	94.37	100.00	100.00	94.84	94.84

Table 3. Performances comparison on the LLCM dataset by different methods

View table

Table 3. Performances comparison on the LLCM dataset by different methods

Method	Rank-1 accuracy		mAP
Method	Visible to infrared	Infrared to visible	Visible to infrared	Infrared to visible
DDAG^［17］	40.3	48.0	48.4	52.3
AGW^［14］	43.6	51.5	51.8	55.3
MMN^［31］	52.5	59.9	58.9	62.7
DEEN^［13］	54.9	62.5	62.9	65.8
IDKL^［26］	72.22	70.72	66.43	65.19
Proposed	73.63	71.22	65.83	65.70

Table 4. Ablation experimental results of different components on the SYSU-MM01 dataset under all-search and single-shot setting unit: %

View table

Table 4. Ablation experimental results of different components on the SYSU-MM01 dataset under all-search and single-shot setting unit: %

Method	Baseline	TGSA+CSA	L_s	L_pr	FCA	L_var	Rank-1 accuracy	mAP
a	√						65.81	66.49
b	√				√		67.18	67.59
c	√	√					69.88	71.69
d	√	√	√				77.51	76.80
e	√	√	√		√		78.28	77.80
f	√	√	√	√			79.01	78.79
g	√	√	√	√		√	80.24	79.89
h	√	√	√	√	√		80.87	80.26
i	√	√	√	√	√	√	82.14	81.59

Table 5. Comparison of FCA and other attention modules on the LLCM dataset

View table

Table 5. Comparison of FCA and other attention modules on the LLCM dataset

Method	Rank-1 accuracy		mAP
Method	Visible to infrared	Infrared to visible	Visible to infrared	Infrared to visible
FCA→NL^［33］	67.89	65.32	56.57	55.63
FCA→SNT^［34］	68.84	66.54	59.22	57.66
FCA→SE^［27］	70.01	69.27	60.71	63.19
FCA→CBAM^［35］	72.94	71.01	63.46	64.88
Proposed	73.63	71.22	65.83	65.70

Table 6. Model complexity results
View table
Table 6. Model complexity results
Model Model memory /MB Training time /s
AGW^［14］ 272.22 92.14
DEEN^［13］ 339.66 284.90
Proposed 297.73 173.21

Tools

Get Citation

Copy Citation Text

Taizhe Tan, Mengrou Li, Zhuo Yang, Zhiyuan Gong. Cross-Modal Pedestrian Re-Identification Combining Frequency-Domain Attention and Modal Co-Feature Optimization[J]. Laser & Optoelectronics Progress, 2025, 62(12): 1215010

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites