MLLMs-Augmented Visual-Language Representation Learning
This work addresses the need for richer multi-modal data in AI by enhancing visual-language representation learning, offering an incremental improvement through MLLM-augmented captioning.
The paper tackles the problem of limited image-text associations in visual-language pre-training by using Multi-modal Large Language Models (MLLMs) to generate multiple diverse captions for images, with a 'text shearing' method to mitigate hallucinations and monotony, resulting in improvements of 5.6 to 35.0 and 16.8 to 46.1 in Recall@1 for fine-tuning and zero-shot image-text retrieval, respectively.
Visual-language pre-training has achieved remarkable success in many multi-modal tasks, largely attributed to the availability of large-scale image-text datasets. In this work, we demonstrate that Multi-modal Large Language Models (MLLMs) can enhance visual-language representation learning by establishing richer image-text associations for image-text datasets. Our approach is simple, utilizing MLLMs to extend multiple diverse captions for each image. To prevent the bias introduced by MLLMs' hallucinations and monotonous language styles, we propose "text shearing" to maintain the quality and availability of extended captions. In image-text retrieval, without introducing additional training cost, our method consistently obtains 5.6 ~ 35.0 and 16.8 ~ 46.1 improvement on Recall@1 under the fine-tuning and zero-shot settings, respectively. Notably, we obtain zero-shot results that are comparable to fine-tuning on target datasets, which encourages more exploration of the versatile use of MLLMs.