This paper addresses the task of generating fluent descriptions by training on a non-uniform combination of data sources, containing both human-annotated and web-collected captions. Large-scale datasets with noisy image-text pairs, indeed, provide a sub-optimal source of supervision because of their low-quality descriptive style, while human-annotated datasets are cleaner but smaller in scale. To get the best of both worlds, we propose to leverage and separate semantics and descriptive style through the incorporation of a style token and keywords extracted through a retrieval component. The proposed model avoids the need of object detectors, is trained with a single objective of prompt language modeling, and can replicate the style of human-collected captions while training on sources with different input styles. Experimentally, the model shows a strong capability of recognizing real-world concepts and producing high-quality captions. Extensive experiments are performed on different image captioning datasets, including CC3M, nocaps, and the competitive COCO dataset, where our model consistently outperforms baselines and state-of-the-art approaches.
View on arXiv