VideoOFA: Two-Stage Pre-Training for Video-to-Text Generation

TL;DR

VideoOFA introduces a two-stage pre-training approach for video-to-text tasks, leveraging image-text data and video-specific training to significantly improve performance on captioning and question answering benchmarks.

Contribution

The paper presents a novel two-stage pre-training framework that enhances video-to-text generation by combining image-text pre-training with video-specific adaptation.

Findings

Achieves state-of-the-art results on four video captioning benchmarks.

Outperforms existing models on two video question answering datasets.

Demonstrates strong generalization as a universal video-to-text model.

Abstract

We propose a new two-stage pre-training framework for video-to-text generation tasks such as video captioning and video question answering: A generative encoder-decoder model is first jointly pre-trained on massive image-text data to learn fundamental vision-language concepts, and then adapted to video data in an intermediate video-text pre-training stage to learn video-specific skills such as spatio-temporal reasoning. As a result, our VideoOFA model achieves new state-of-the-art performance on four Video Captioning benchmarks, beating prior art by an average of 9.7 points in CIDEr score. It also outperforms existing models on two open-ended Video Question Answering datasets, showcasing its generalization capability as a universal video-to-text model.