CFIR: Fast and Effective Long-Text To Image Retrieval for Large Corpora

Text-to-image retrieval aims to find the relevant images based on a text query, which is important in various use-cases, such as digital libraries, e-commerce, and multimedia databases. Although Multimodal Large Language Models (MLLMs) demonstrate state-of-the-art performance, they exhibit limitations in handling large-scale, diverse, and ambiguous real-world needs of retrieval, due to the computation cost and the injective embeddings they produce. This paper presents a two-stage Coarse-to-Fine Index-shared Retrieval (CFIR) framework, designed for fast and effective large-scale long-text to image retrieval. The first stage, Entity-based Ranking (ER), adapts to long-text query ambiguity by employing a multiple-queries-to-multiple-targets paradigm, facilitating candidate filtering for the next stage. The second stage, Summary-based Re-ranking (SR), refines these rankings using summarized queries. We also propose a specialized Decoupling-BEiT-3 encoder, optimized for handling ambiguous user needs and both stages, which also enhances computational efficiency through vector-based similarity inference. Evaluation on the AToMiC dataset reveals that CFIR surpasses existing MLLMs by up to 11.06 Recall@1000, while reducing training and retrieval times by 68.75 respectively. We will release our code to facilitate future research at https://github.com/longkukuhi/CFIR.