Universal Captioner: Long-Tail Vision-and-Language Model Training through Content-Style Separation

While captioning models have obtained compelling results in describing natural images, they still do not cover the entire long-tail distribution of real-world concepts. In this paper, we address the task of generating human-like descriptions with in-the-wild concepts by training on web-scale automatically collected datasets. To this end, we propose a model which can exploit noisy image-caption pairs while maintaining the descriptive style of traditional human-annotated datasets like COCO. Our model separates content from style through the usage of keywords and stylistic tokens, employing a single objective of prompt language modeling and being simpler than other recent proposals. Experimentally, our model consistently outperforms existing methods in terms of caption quality and capability of describing long-tail concepts, also in zero-shot settings. According to the CIDEr metric, we obtain a new state of the art on both COCO and nocaps when using external data.