Researchers introduce FSIR benchmark dataset to improve AI image retrieval from text using few-shot learning techniques.

New Few-Shot Text-to-Image Retrieval (FSIR) task addresses limitations of pre-trained vision-language models in handling compositional and out-of-distribution queries

FSIR-BD benchmark dataset contains 38,353 images and 303 queries, with 82% focusing on challenging compositional scenarios across urban scenes and nature species

Few-shot learning approach enables models to learn from minimal examples, mimicking human cognitive abilities for improved image retrieval performance

Current VLM embedding-based retrieval struggles with complex compositional queries and out-of-distribution image-text pairs, which the new dataset explicitly targets