Vision-Language Pretraining

Vision-language pretraining teaches models to connect visual content with language, supporting image-text retrieval, zero-shot classification, question answering, and applications such as medical image analysis. Common approaches include contrastive image-text alignment, caption generation, masked or predictive feature learning, and region-level grounding, trained on web captions, curated datasets, or domain-specific reports. Researchers evaluate transfer on classification and retrieval benchmarks such as ImageNet and COCO, as well as on dense perception, visual question answering, documents, and clinical tasks. A central tension is that global alignment scales well but can miss fine details and composition, while local or generative objectives offer richer supervision but vary in benefit across tasks and settings.