VGSG: Vision-Guided Semantic-Group Network for Text-Based Person Search

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Text-based Person Search (TBPS) requires aligning fine-grained visual and textual features across modalities.
• Existing TBPS methods often rely on inefficient external tools or complex cross-modal interactions for feature alignment.

Purpose of the Study:

• To propose an efficient Vision-Guided Semantic-Group Network (VGSG) for text-based person search.
• To extract well-aligned fine-grained visual and textual features without external tools or heavy cross-modal interaction.

Main Methods:

• Developed a Semantic-Group Textual Learning (SGTL) module to group textual features based on semantic cues.
• Implemented a Vision-guided Knowledge Transfer (VGKT) module using vision-guided attention for feature alignment.
• Introduced relational knowledge transfer (vision-language similarity and class probability transfer) for adaptive information propagation.

Main Results:

• The VGSG network effectively extracts aligned fine-grained visual and textual features.
• SGTL implicitly groups similar semantic patterns in textual features.
• VGKT aligns semantic-group textual features with visual features using relational knowledge transfer.

Conclusions:

• The proposed VGSG method achieves superior performance on text-based person search benchmarks.
• The approach offers an efficient and effective solution for cross-modal feature alignment in TBPS.