OKVQA

OK-VQA是用于视觉问题回答的新数据集，需要可以利用外部知识来回答问题的方法。
14,055开放式问题
每个问题5个地面真相答案
手动过滤以确保所有问题都需要外部知识 (例如来自Wikipeida)
减少最常见答案的问题，以减少数据集偏差

OK-VQA is a novel visual question answering (VQA) dataset that requires methods capable of leveraging external knowledge to answer questions. It includes the following key features:
1. 14,055 open-ended questions
2. Five ground-truth answers per question
3. Manually filtered to ensure all questions necessitate external knowledge, such as that sourced from Wikipedia
4. Downsampling of questions with overly frequent standard answers to mitigate dataset bias

搜集汇总

背景与挑战

背景概述

OKVQA是一个视觉问题回答数据集，包含14,055个需要外部知识回答的开放式问题，每个问题提供5个标准答案，旨在减少数据集偏差并促进利用外部知识的视觉问答方法研究。