Q-GrAM: Fine-Grained Image-Text Retrieval via Grouped Query Routing and Conditional Query Modulation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current image-text retrieval methods use global vectors, limiting fine-grained semantic understanding.
• Textual queries with objects, attributes, and relations are challenging for existing models.

Purpose of the Study:

• To propose Q-GrAM, an adaptation of BLIP-2 Q-Former for fine-grained text-to-image retrieval.
• To enhance the retrieval of images based on detailed textual descriptions.

Main Methods:

• Q-GrAM partitions query budget into semantically distinct groups.
• A text-guided router assigns semantic demands, and conditional initialization modulates groups.
• Group-aware late interaction scoring matches visual and textual tokens.

Main Results:

• Q-GrAM demonstrates strong performance on MS-COCO 5K, Flickr30K, and Flickr30K-CFQ datasets.
• Achieves superior text-to-image retrieval compared to global embedding and fine-grained matching methods.
• Maintains competitive bidirectional retrieval performance.

Conclusions:

• Structured, text-conditioned query specialization is effective for fine-grained retrieval.
• Q-GrAM advances the capabilities of text-driven image search systems.