Effective End-to-End Vision Language Pretraining with Semantic Visual Loss

TL;DR

This paper proposes a new end-to-end vision language pretraining approach using auxiliary visual losses, achieving faster training, improved accuracy, and comparable or better performance than region-based models, with significantly faster inference.

Contribution

Introduces three auxiliary visual losses to enhance end-to-end vision language pretraining, enabling faster convergence and better downstream performance with reduced pretraining resources.

Findings

End-to-end models match or outperform region-based models on downstream tasks.

Models run over 10 times faster during inference.

Pretraining with only 10% of GPU hours achieves comparable results.

Abstract

Current vision language pretraining models are dominated by methods using region visual features extracted from object detectors. Given their good performance, the extract-then-process pipeline significantly restricts the inference speed and therefore limits their real-world use cases. However, training vision language models from raw image pixels is difficult, as the raw image pixels give much less prior knowledge than region features. In this paper, we systematically study how to leverage auxiliary visual pretraining tasks to help training end-to-end vision language models. We introduce three types of visual losses that enable much faster convergence and better finetuning accuracy. Compared with region feature models, our end-to-end models could achieve similar or better performance on downstream tasks and run more than 10 times faster during inference. Compared with other end-to-end models, our proposed method could achieve similar or better performance when pretrained for only 10% of the pretraining GPU hours.