Vision-Language Retrieval

Vision-language retrieval studies how to rank images, video, and visual documents against text queries, from short captions to detailed or compositional descriptions. Common approaches include CLIP-style contrastive dual encoders, fine-grained attribute or region alignment, and coarse-to-fine pipelines that rerank candidates with vision-language models or language-model reasoning. Benchmarks span image-text datasets such as COCO and Flickr, video retrieval, and specialized settings including visual documents, geolocation, and fine-grained product search. A central tension is that fast similarity scores can miss logical constraints or subtle evidence, while richer reranking and reasoning improve precision at greater computational cost and may generalize unevenly across domains.