Learning Weak Semantics by Feature Graph for Attribute-Based Person Search.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Attribute-based person search is crucial for surveillance when visual data is unreliable.
• Current datasets suffer from attribute labeling problems, causing semantic gaps between text and images.
• Existing methods often overlook these data imperfections, limiting performance.

Purpose of the Study:

• To address attribute labeling issues in person search datasets.
• To reduce the semantic gap in text-to-image retrieval for person identification.
• To enhance the representability of attribute features and improve search accuracy.

Main Methods:

• Proposed Weak Semantic Embeddings (WSEs) to modify attribute text data distribution.
• Introduced feature graphs to learn collaborative and calibrated information.
• Modeled relationships between semantic embeddings using feature graphs to bridge the text-image semantic gap.

Main Results:

• Extensive evaluations on PETA, Market-1501 Attribute, and PA100K benchmarks.
• Demonstrated the effectiveness of the proposed WSEs in improving attribute feature representation.
• Achieved state-of-the-art performance, outperforming existing methods on challenging datasets.

Conclusions:

• The proposed WSEs effectively address attribute labeling problems in person search.
• Feature graphs enhance information learning and reduce the semantic gap in text-to-image retrieval.
• The method significantly improves attribute-based person search accuracy and robustness.