Address the Unseen Relationships: Attribute Correlations in Text Attribute Person Search

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Text attribute person search identifies pedestrians using textual descriptions, crucial when image samples are unavailable.
• Existing methods often overlook latent correlations between attributes, limiting performance.
• Developing robust methods for text-based person identification is essential for surveillance and security.

Purpose of the Study:

• To propose a novel graph convolutional network (GCN) and pseudo-label-based method for text attribute person search.
• To effectively model and leverage latent correlations between textual attributes.
• To improve the accuracy and robustness of person identification from textual descriptions.

Main Methods:

• Constructing attribute correlations using label co-occurrence probability via a GCN.
• Representing attributes as nodes and their correlations as edges in a graph.
• Combining a cross-attention module (CAM) with GCN for enhanced representations.
• Utilizing pseudo-labeling on the test set to adapt to unseen attribute relationships.

Main Results:

• The proposed GCN-based method significantly outperforms state-of-the-art approaches.
• Effective modeling of attribute correlations leads to improved person search accuracy.
• The model demonstrates robustness in handling unseen attribute relationships.

Conclusions:

• The developed GCN and pseudo-labeling approach offers a significant advancement in text attribute person search.
• Leveraging attribute correlations and adaptive learning enhances identification performance.
• This method provides a more practical solution for person identification scenarios relying solely on witness accounts.