Multimodal Reranking

Multimodal reranking studies how to improve an initial retrieval shortlist by judging how well each candidate matches a query across images, text, video, speech, or other evidence. Approaches include vision-language model scoring, pairwise comparisons and listwise ranking, cross-modal embedding methods, and structured evidence such as regions, attributes, or semantic constraints. These methods are used in settings ranging from image and video search to document retrieval, visual question answering, and person re-identification, with benchmarks measuring top-rank accuracy, recall, and ranking quality. A central tension is that detailed multimodal reasoning can improve fine-grained relevance judgments but is costly and may be poorly calibrated, motivating more efficient, selective, and reliable reranking methods.