.png){kind=icon}
.png){kind=icon}
Return
A domain generalization pedestrian re-identification algorithm based on meta-graph aware
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-023-15765-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Domain generalization is a key problem to solve the difference between the source domain and the target domain. This paper proposes a person re-identification algorithm based on meta-graph aware (Meta-GA) under the framework of meta-learning, which includes two stages: meta-global aware (M-GA) and meta-graph relationship sampling (M-GRS). In order to reduce inter-domain differences, a meta-global aware mechanism is proposed to construct an interaction model (paired relationship) in the meta training domain by stacking affinity models and dividing saliency features between the pedestrians. Then a learning interaction model is used to construct a global knowledge map to classify and weighted the structural information. In order to accurately learn the discriminative features, a meta-graph relationship sampling model is proposed. The similarity of the pedestrian cross-domain features between the domains is used to construct a feature relationship map between the adjacent classes. To enhance domain invariant features and improve the model generalization, positive samples are sampled cyclically and negative samples are sampled randomly. On this basis, the gradient norm is trimmed to prevent the model overfitting. The experimental results show that the robustness and accuracy of the proposed algorithm have been significantly improved. In the Market-1501 to DukeMTMC-ReID experiment, Rank-1 and mAP increased by 5.25% and 3.73%, respectively. In the DukeMTMC-ReID to Market-1501 experiment, Rank-1 and mAP increased by 1.73% and 0.93%, respectively, which are significantly superior to those of the recent representative algorithms.
Keywords:
Person re-identification
Domain generalization
Meta-Learning
Graph Aware
Inter-domain differences
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W