We train a multi-modal embedding to associate fragments of images (objects) and sentences (noun and verb phrases) with a structured, max-margin objective. Our model enables efficient and interpretible retrieval of images from sentence descriptions (and vice versa).
