E2E-VLP: End-to-End Vision-Language Pre-training Enhanced by Visual Learning

TL;DR

E2E-VLP introduces an end-to-end vision-language pre-training framework that learns pixel-level visual representations and cross-modal alignments within a unified Transformer encoder-decoder. By incorporating DETR-inspired object detection and image-caption generation as joint pre-training tasks, it achieves competitive V+L performance with fewer parameters and significantly faster inference than region-based two-stage models. The approach demonstrates that end-to-end grid features, when guided by fine-grained visual tasks, can match or surpass traditional detectors in cross-modal understanding and generation. The work highlights the practicality of end-to-end VLP for scalable, efficient vision-language tasks and sets a direction for deeper bottom-layer fusion and expanded pre-training objectives.

Abstract

Vision-language pre-training (VLP) on large-scale image-text pairs has achieved huge success for the cross-modal downstream tasks. The most existing pre-training methods mainly adopt a two-step training procedure, which firstly employs a pre-trained object detector to extract region-based visual features, then concatenates the image representation and text embedding as the input of Transformer to train. However, these methods face problems of using task-specific visual representation of the specific object detector for generic cross-modal understanding, and the computation inefficiency of two-stage pipeline. In this paper, we propose the first end-to-end vision-language pre-trained model for both V+L understanding and generation, namely E2E-VLP, where we build a unified Transformer framework to jointly learn visual representation, and semantic alignments between image and text. We incorporate the tasks of object detection and image captioning into pre-training with a unified Transformer encoder-decoder architecture for enhancing visual learning. An extensive set of experiments have been conducted on well-established vision-language downstream tasks to demonstrate the effectiveness of this novel VLP paradigm.

Figures (1)

• Figure 1: The overall framework of E2E-VLP. Our model employs a unified encoder-decoder transformer framework to learn visual representation, and semantic alignment between image and text jointly.