Hybrid Attention Network for Language-Based Person Search.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Language-based person search is a complex fine-grained cross-modal retrieval task.
• Existing methods face challenges in accurately matching natural language descriptions to person images.

Purpose of the Study:

• To propose a novel hybrid attention network for enhanced language-based person search.
• To improve the discriminative feature representation of person images and language descriptions.
• To effectively address cross-modal heterogeneity in retrieval.

Main Methods:

• A cubic attention mechanism combining cross-layer spatial and channel attention for person images.
• A text attention network using bidirectional LSTM (BiLSTM) and self-attention for language descriptions.
• A cross-modal attention mechanism and joint loss function for improved cross-modal learning.

Main Results:

• The proposed hybrid attention network achieved higher performance on the CUHK-PEDES dataset.
• Demonstrated superior fine-grained feature representation for person images.
• Showcased enhanced capability in capturing semantic dependencies in language descriptions.

Conclusions:

• The novel hybrid attention network significantly advances language-based person search.
• The approach effectively leverages both intra-modal and cross-modal correlations.
• This method offers a promising solution for fine-grained cross-modal retrieval tasks.