Organizations: Faculty of Data Science, City University of Macau, Macau SAR, China · School of Big Data, Zhuhai College of Science and Technology, Zhuhai 519041, China · Faculty of Innovation Engineering, Macau University of Science and Technology, Macau SAR, China · School of Computer Science and Technology, Anhui University, Hefei 230601, China
Pedestrian Attribute Recognition is a foundational computer vision task that provides essential support for downstream applications, including person retrieval in video surveillance and intelligent retail analytics. However, existing research is frequently constrained by the ``one-model-per-dataset" paradigm and struggles to handle significant discrepancies across domains in terms of modalities, attribute definitions, and environmental scenarios. To address these challenges, we propose UniPAR, a unified Transformer-based framework for PAR. By incorporating a unified data scheduling strategy and a dynamic classification head, UniPAR enables a single model to simultaneously process diverse datasets from heterogeneous modalities, including RGB images, video sequences, and event streams. We also introduce an innovative phased fusion encoder that explicitly aligns visual features with textual attribute queries through a late deep fusion strategy. Experimental results on the widely used benchmark datasets, including MSP60K, DukeMTMC, and EventPAR, demonstrate that UniPAR achieves performance comparable to specialized SOTA methods. Furthermore, multi-dataset joint training significantly enhances the model's cross-domain generalization and recognition robustness in extreme environments characterized by low light and motion blur. The source code of this paper will be released on https://github.com/Event-AHU/OpenPAR
Figures & tables
Figure 1: Comparison between existing PAR models and our newly proposed one.
Figure 2: The overall architecture of our proposed UnifiedPAR framework
Dataset
mA
Accuracy
Precision
Recall
F1
Individually Trained
MSP60K [ 12 ]
75.12
76.32
86.77
84.32
85.15
DUKE [ 24 ]
69.73
61.76
75.58
73.78
74.09
EventPAR [ 34 ]
86.90
82.31
87.01
88.41
87.53
Jointly Trained
MSP60K [ 12 ]
79.55
77.79
85.86
87.54
86.32
DUKE [ 24 ]
75.56
70.40
82.77
79.92
80.75
EventPAR [ 34 ]
88.51
85.21
89.22
89.77
89.36
Table 1: The experimental result of our proposed method.
Methods
Publish
Code
mA
Acc
Prec
Recall
F1
#01 DeepMAR [ 14 ]
ACPR15
URL
70.46
72.83
84.71
81.46
83.06
#02 Strong Baseline [ 9 ]
arXiv20
URL
74.09
73.74
84.06
83.51
83.31
#03 RethinkingPAR [ 9 ]
arXiv20
URL
74.01
74.20
84.17
83.94
84.06
#04 SSCNet [ 10 ]
ICCV21
URL
69.71
69.31
79.22
82.47
80.82
#05 VTB [ 3 ]
TCSVT22
URL
76.09
75.36
83.56
86.46
84.56
#06 Label2Label [ 18 ]
ECCV22
URL
73.61
72.66
81.79
84.32
82.56
Table 2: Comparison with state-of-the-art methods on MSP60K datasets.
Methods
Publish
Code
mA
Acc
Prec
Recall
F1
#01 DeepMAR [ 14 ]
ACPR15
URL
66.57
69.53
74.90
89.54
81.57
#02 ALM [ 26 ]
ICCV19
URL
57.18
64.17
75.59
73.20
74.38
#03 Strong Baseline [ 9 ]
arXiv20
URL
73.75
61.86
67.23
80.78
75.43
#04 RethinkingPAR [ 9 ]
arXiv20
URL
81.37
80.84
86.31
87.57
86.93
#05 SSCNet [ 10 ]
ICCV21
URL
63.10
66.07
72.72
83.22
77.62
#06 VTB [ 3 ]
TCSVT22
URL
88.41
83.83
87.89
89.51
88.53
Table 3: Comparison with state-of-the-art methods on EventPAR dataset.
Jointly Trained
mA
Accuracy
Precision
Recall
F1 score
MSP60k
80.06
77.65
85.51
87.63
86.18
DUKE
73.06
68.18
81.04
78.79
79.24
MSP60k
79.34
77.01
84.90
87.52
85.80
EventPAR
88.78
85.20
89.42
89.55
89.35
DUKE
74.13
66.38
79.10
77.54
77.68
EventPAR
88.87
83.47
87.78
89.15
88.30
Table 4: Results of Cross-dataset Evaluation on MSP60k, DUKE, and EventPAR.
Dataset
mA
Accuracy
Precision
Recall
F1
Without Attribute Set
MSP60K [ 12 ]
77.33
71.51
80.09
84.64
81.84
DUKE [ 24 ]
69.04
59.26
75.71
71.13
72.26
EventPAR [ 34 ]
86.56
81.73
86.60
88.14
87.19
BERT Embedding
MSP60K [ 12 ]
79.22
77.11
85.37
87.11
85.84
DUKE [ 24 ]
72.94
66.87
80.29
77.57
78.22
EventPAR [ 34 ]
88.47
85.05
89.20
89.61
89.26
Table 5: The Ablation Study Result on Text Encoder.
Lossrate
Dataset
mA
Accuracy
Precision
Recall
F1 score
1,1,1
MSP60k
78.85
76.31
84.36
87.13
85.31
DUKE
72.39
66.32
80.11
76.90
77.76
EventPAR
88.24
85.08
89.20
89.63
89.28
0.8,1,1
MSP60k
76.75
73.54
81.94
85.88
83.40
DUKE
71.02
64.53
78.01
76.46
76.46
EventPAR
87.65
84.20
88.84
89.93
88.74
Table 6: Results of Parameter Analysis.
Figure 3: Visual Analysis of the Generalization Performance of Our Proposed Model on the MSP60k, DUKE, and EventPAR Datasets. Red markers indicate misclassified attributes, while orange markers denote undetected attributes.
Figure 4: Visualizing the Cross-Dataset Generalization of Our Proposed Model. Red markers indicate misclassified attributes, while orange markers denote undetected attributes.
City University of Macau, Macau SAR, China · Zhuhai College of Science and Technology, Zhuhai, China · Macau University of Science and Technology, Macau SAR, China +1