HAVEN
收藏资源简介:
HAVEN是一个用于评估大型多模态模型在视频理解中的幻觉现象的基准数据集。该数据集基于三个维度构建:幻觉的原因(与先验知识冲突、上下文冲突、模型能力不足)、幻觉的方面(对象、场景、事件)以及问题的格式(二选一、多选、简答)。数据集包含来自公开视频数据集(如COIN、ActivityNet和Sports1M)以及从互联网手动收集的视频片段。
HAVEN is a benchmark dataset designed to evaluate the illusion phenomenon in video understanding for large-scale multimodal models. The dataset is constructed based on three dimensions: the cause of illusion (conflict with prior knowledge, context conflict, insufficient model capability), the aspect of illusion (object, scene, event), and the format of the question (choose one, multiple choice, short answer). The dataset includes video clips from public video datasets (such as COIN, ActivityNet, and Sports1M) as well as manually collected videos from the internet.
HAVEN 数据集概述
📌 数据集简介
- 名称:HAVEN (HAllucination in Video UndErstaNding)
- 目的:定量评估大型多模态模型(LMMs)在视频理解中的幻觉现象
- 特点:专注于视频理解中的幻觉问题,区别于传统的图像理解基准
📊 数据构成
- 视频来源:
- 公开数据集:COIN、ActivityNet、Sports1M
- 手动收集:来自互联网的视频片段
- 样本量:包含1,200个问题
🔍 数据维度
- 幻觉原因:
- 与先验知识冲突
- 上下文冲突
- 模型固有能力缺陷
- 幻觉方面:
- 对象
- 场景
- 事件
- 问题格式:
- 二元选择
- 多项选择
- 简答
📂 数据结构
json { "Index": 1, "Question": "Are the individuals in the video sewing garments with thread?", "Answer": "No", "Causes": "Conflict with prior", "Aspects": "Event", "Form": "Binary-choice", "Source Link": "https://www.youtube.com/embed/xZecGPPhbHE", "Begin": "0:29", "End": "0:40", "Video Path": "Coin0001", "Group_id": "Coin_group0001" }
📝 字段说明
- Index:全局问题索引
- Causes:幻觉原因分类
- Aspects:视频内容中的幻觉类别
- Form:问题格式
- Source Link:原始视频来源
- Begin/End:相关视频片段时间戳
- Video Path:视频唯一标识符
- Group_id:相关问题组标识
🛠 使用工具
- Infer.py:在数据集上运行模型
- Judge.py:使用GPT-4o-mini评估模型预测
📄 相关文献
- 论文标题:Exploring Hallucination of Large Multimodal Models in Video Understanding: Benchmark, Analysis and Mitigation
- arXiv链接:https://arxiv.org/abs/2503.19622




