EgoExoBench
收藏资源简介:
EgoExoBench是一个针对多模态大型语言模型(MLLMs)的第一人称(自我中心)和第三人称(外部中心)视频理解和推理的基准。该数据集由11个子任务组成,涵盖了超过7,300个问答对,分为三个核心挑战:语义对齐、视角关联和时间推理。这些数据来自于公开可用的数据集,经过严格的注释协议整理。EgoExoBench旨在评估模型在自我中心和外中心视角之间整合信息的能力,为研究具身代理和智能助手寻求类似人类的跨视角智能提供宝贵的资源。
EgoExoBench is a benchmark for first-person (egocentric) and third-person (exocentric) video understanding and reasoning for multimodal large language models (MLLMs). This dataset comprises 11 subtasks covering over 7,300 question-answer pairs, and falls under three core challenge categories: semantic alignment, perspective association, and temporal reasoning. All data are sourced from publicly available datasets and curated following strict annotation protocols. EgoExoBench is designed to evaluate models' capability to integrate information across egocentric and exocentric perspectives, serving as a valuable resource for research on embodied agents and intelligent assistants targeting human-like cross-perspective intelligence.
EgoExoBench 数据集概述
基准测试概述
- 目的:评估多模态大语言模型(MLLMs)在跨视角视频理解中的表现。
- 规模:包含超过7,300个多选题,覆盖11个子任务。
- 关键维度:
- Ego-Exo Relation
- Ego-Exo View Transition
- Ego-Exo Temporal Reasoning
数据来源
数据准备
-
目录结构:
EgoExoBench/ └── data/ ├── CVMHAT/ │ └── data ├── Ego-Exo4D/ │ └── takes/ ├── EgoExoLearn/ ├── EgoMe/ ├── LEMMA/ └── TF2023/ └── data/
-
数据处理:
- 对CVMHAT和TF2023数据集使用边界框注释增强原始帧。
- 运行命令: shell python data/CVMHAT/tools/process_bbox.py python data/TF2023/tools/process_bbox.py
-
多选题下载:从指定链接下载并放置于
MCQ/目录。
安装与评估
-
安装: shell git clone https://github.com/ayiyayi/EgoExoBench.git cd EgoExoBench
-
评估:基于VLMEvalKit,运行命令示例: shell torchrun --nproc-per-node=1 run.py --data EgoExoBench_MCQ --model Qwen2.5-VL-7B-Instruct-ForVideo python run.py --data EgoExoBench_MCQ --model Qwen2.5-VL-72B-Instruct-ForVideo

- 1EgoExoBench: A Benchmark for First- and Third-person View Video Understanding in MLLMs上海人工智能实验室, 南京大学, 东京大学, 浙江大学, 复旦大学 · 2025年



