VL-BEiT: Generative Vision-Language Pretraining

TL;DR

VL-BEiT is a unified vision-language Transformer pretraining approach that uses masked prediction tasks on images and texts, achieving strong results across multiple vision-language benchmarks and transferable visual tasks.

Contribution

It introduces a simple, unified generative pretraining framework for vision-language models with a shared Transformer backbone trained from scratch.

Findings

Strong performance on vision-language tasks like VQA, reasoning, and retrieval.

Transferable visual features for classification and segmentation.

Effective one-stage training with a shared model for multiple tasks.

Abstract

We introduce a vision-language foundation model called VL-BEiT, which is a bidirectional multimodal Transformer learned by generative pretraining. Our minimalist solution conducts masked prediction on both monomodal and multimodal data with a shared Transformer. Specifically, we perform masked vision-language modeling on image-text pairs, masked language modeling on texts, and masked image modeling on images. VL-BEiT is learned from scratch with one unified pretraining task, one shared backbone, and one-stage training. Our method is conceptually simple and empirically effective. Experimental results show that VL-BEiT obtains strong results on various vision-language benchmarks, such as visual question answering, visual reasoning, and image-text retrieval. Moreover, our method learns transferable visual features, achieving competitive performance on image classification, and semantic segmentation.