I-VQA 数据集
收藏资源简介:
I-VQA 数据集是一种新型的视频问答数据集,旨在解决视频问答中无法直接定位显式视觉证据的问题。该数据集通过对 Grounded VideoQA 数据集中的视频进行重新组织和去除显式视觉证据片段,以支持隐式视频问答任务的训练和评估。数据集包含 5549 个问答样本,其中 3749 个样本用于训练集,1800 个样本分为测试集。数据集中的问题主要关注于视频中的隐含意义和深层意图,而不仅仅是显式的视觉证据。I-VQA 数据集的创建为隐式视频问答任务的研究提供了重要的数据基础,有助于推动该领域的发展。
The I-VQA dataset is a novel video question answering (VideoQA) dataset aimed at solving the problem that existing VideoQA tasks cannot directly locate explicit visual evidence. This dataset is constructed by reorganizing the videos from the Grounded VideoQA dataset and removing segments containing explicit visual evidence, so as to support the training and evaluation of implicit video question answering tasks. It comprises 5,549 question-answering (QA) pairs, among which 3,749 pairs are assigned to the training set, and the remaining 1,800 pairs are allocated to the test set. The questions in this dataset mainly focus on the implied meanings and deep intentions within videos, rather than just explicit visual evidence. The creation of the I-VQA dataset provides an important data foundation for research on implicit video question answering tasks, and helps promote the development of this field.
I-VQA 数据集概述
数据集基本信息
- 名称: I-VQA Dataset
- 样本数量: 5,549个QA样本
- 训练集: 3,749个样本
- 测试集: 1,800个样本
数据标注
- 训练集标注文件:
dataset/NExT-GQA/ICR/training_4k.json - 测试集标注文件:
dataset/NExT-GQA/ICR/test_mc.json - 视频来源: Next-GQA, E.T. Bench, REXTIME(根据提供的视频ID查找)
标注内容详情
- duration: 显式证据(在I-VQA任务中被屏蔽)
- context: 上下文问答对
- action_relation_intent: 由GPT-4生成的线索
- context_relation: 上下文线索与当前隐式问题的关系(1表示相关,0表示无关)
相关文件
- SUTD-Traffic QA样本:
dataset/SUTD-traffic/filtered_data2.json - PSAV样本:
dataset/ads/converted_annotation.json - IRM测试结果:
- 多选:
Implicit_reasoner/result/IRM_test_mc.json - 开放式:
Implicit_reasoner/result/IRM_test_open.json
- 多选:
- GPT评分/评判脚本:
Implicit_reasoner/eval_GPT_score.py




