Overview - ICDAR 2023 Competition on Text-based Video Question Answering on News Videos

Introduction

Although challenges with video comprehension are not strictly of interest to the Document analysis community, identifying and recognizing text in videos has been an important topic of research within the community. NewsVideoQA challenge aims to promote the task of text-based video question answering and assess current methods on NewsVideoQA. The challenge introduces for the first time text-based video question answering on news videos, which requires systems to analyze the textual content in these videos and use the textual information from multiple frames of the video to provide answers to questions. The past has presented a number of difficulties for text detection, recognition, and tracking text in videos. The last several editions of competitions in ICDAR have seen a rise in community interest in shifting from classic document analysis tasks like detection and recognition to higher-level challenges like question answering on document images, natural scene images containing text, infographics and so on. With the NewsVideoQA challenge, we aim to expand this line of efforts to the video realm. 

 

 

The challenge introduces for the first time text-based video question answering on news videos, which requires systems to analyze the textual content in these videos and use the textual information from multiple frames of the video to provide answers to questions. The NewsVideoQA challenge aims to promote the task of text-aware video question answering and assess current  VQA and VideoQA methods on NewsVideoQA dataset. The NewsVideoQA dataset comprises 10,000 questions framed on 3,083 news video clips.

 

 

References

[1] Soumya Jahagirdar, Minesh Mathew, Dimosthenis Karatzas, C. V. Jawahar, "Watching the News: Towards VideoQA Models that can Read" https://arxiv.org/abs/2211.05588, WACV, 2023