MLLMs-Augmented Visual-Language Representation Learning

TL;DR

This paper leverages Multi-modal Large Language Models to enrich image-text datasets with diverse captions, significantly improving visual-language representation learning and retrieval performance without extra training costs.

Contribution

It introduces a simple method using MLLMs for caption extension and a text shearing technique to improve multi-modal learning, achieving notable zero-shot and fine-tuning improvements.

Findings

Significant improvements in Recall@1 for image-text retrieval.

Zero-shot results comparable to fine-tuning on target datasets.

Effective use of MLLMs to enhance dataset quality without additional training.

Abstract

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.