MLLMs-Augmented Visual-Language Representation Learning

TL;DR

This work tackles noise and bias in image-text pre-training data by leveraging a pool of multimodal large language models to generate diverse extended captions for each image. A text shearing mechanism trims these captions to control hallucinations and repetition while preserving semantic content, and both raw and extended captions are used during training with CLIP/BLIP frameworks. Empirically, the approach yields substantial gains in zero-shot and fine-tuned image-text retrieval, as well as improvements in VQA and captioning across CC3M, CC12M, and YFCC15M, without extra training costs. Analyses show that caption diversity from multiple MLLMs and controlled caption length are key to improved visual-language representations, and the method scales well to large datasets with robust gains over baselines.

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.

Figures (13)

• Figure 1: The performance in (a) and (b) is zero-shot image-text retrieval results on MSCOCO using pretrained BLIP. (c) shows that using a single model to generate captions can easily result in similar text structure and insufficient global considerations.
• Figure 2: The illustration of our method. Different MLLMs jointly construct one-to-many image-text pairs with richer semantic associations.
• Figure 3: Statistics of common nouns in captions generated by MLLMs.
• Figure 4: Analysis of extended captions.
• Figure 5: The influence of the token length, batch size, model numbers, and epochs on the performance of visual-language pre-training.