Pedestrian Attribute Recognition Algorithm Based on Multi-Scale Attention Network

Na Li; Yangyang Wu; Ying Liu; Jin Xing

doi:10.3788/LOP202158.0410025

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 4 >Page 0410025 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 4, 0410025 (2021)

Pedestrian Attribute Recognition Algorithm Based on Multi-Scale Attention Network

Na Li^1、2、*, Yangyang Wu^1、2、*, Ying Liu², and Jin Xing¹

Author Affiliations

¹School of Communication and Information Engineering, Xi'an University of Posts & Telecommunications, Xi'an, Shaanxi 710121, China;

²Key Laboratory of Electronic Information Application Technology for Scene Investigation, Ministry of Public Security, Xi'an, Shaanxi 710121, China

show less

DOI: 10.3788/LOP202158.0410025 Cite this Article Set citation alerts

Na Li, Yangyang Wu, Ying Liu, Jin Xing. Pedestrian Attribute Recognition Algorithm Based on Multi-Scale Attention Network[J]. Laser & Optoelectronics Progress, 2021, 58(4): 0410025 Copy Citation Text

show less

Fig. 1. Structure of the pedestrian attribute recognition network

Download full size

Fig. 2. Flow chart of the feature fusion

Download full size

Fig. 3. Channel attention module

Download full size

Methods	RAP		PA-100K
Methods	mA	F1	mA	F1
Baseline	75.67	78.20	77.28	84.52
Baseline+CA	75.89	78.36	77.35	84.67

Table 1. Verification experiment of the channel attention effectiveness unit: %

Method	RAP		PA-100K
Method	mA	F1	mA	F1
Baseline	75.67	78.20	77.28	84.52
Top-down(addition)	74.23	78.01	76.65	84.63
Top-down(weight)	76.75	78.96	78.13	84.95
Top-down(weight)+CA	78.70	80.12	79.82	85.71

Table 2. Recognition results of different feature fusion modules unit: %

Feature layer	Age less 16	Age 17-60	Age bigger 60	ub-shirt	lb-skirt	ub-short sleeve	mA
conv2	49.54	49.36	50.12	51.23	50.56	54.57	61.29
conv3	50.12	49.78	48.93	50.54	54.18	69.55	65.46
conv4	49.82	49.70	47.65	49.73	53.90	79.78	62.91
p5'	56.76	52.59	60.99	71.93	75.68	77.21	77.01
p4'	61.24	54.42	64.23	78.24	74.76	79.92	77.23
p3'	62.35	56.74	67.42	78.89	77.15	80.54	76.89
p2'	62.27	56.67	66.37	77.97	75.43	80.35	77.65
Baseline	63.36	58.49	71.17	78.52	78.82	79.15	75.67
Ours	76.42	75.56	66.92	79.14	80.14	78.66	78.70

Table 3. Recognition results for each layer feature on RAP data set unit: %

Algorithm	mA	Acc	Prec	Rec	F1
ACN	69.66	62.61	80.12	72.26	75.98
DeepMar	73.79	62.02	74.92	76.21	75.56
HP-Net	76.12	65.39	77.33	78.79	78.05
VeSPA	77.70	67.35	79.51	79.67	79.59
PGDM	74.31	64.57	78.86	75.90	77.35
IA2-Net	77.44	65.75	79.01	77.45	78.03
Ours	78.70	68.17	78.89	79.98	80.12

Table 4. Recognition results of different algorithms on the RAP data set unit: %

Algorithm	mA	Acc	Prec	Rec	F1
DeepMar	72.70	70.39	82.24	80.42	81.32
HP-Net	74.21	72.19	82.97	82.09	82.53
VeSPA	76.32	73.00	84.99	81.49	83.20
PGDM	74.95	73.08	84.36	82.24	83.29
IA2-Net	77.28	74.73	83.34	85.73	84.52
Ours	79.82	78.17	82.83	84.98	85.71

Table 5. Recognition results of different algorithms on the PA-100K data set unit: %

Na Li, Yangyang Wu, Ying Liu, Jin Xing. Pedestrian Attribute Recognition Algorithm Based on Multi-Scale Attention Network[J]. Laser & Optoelectronics Progress, 2021, 58(4): 0410025

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information