EgoTaskQA

该数据集名为EgoTaskQA，它为理解人类任务提供了一个基准，通过在真实世界的第一人称视角视频上进行问答，旨在探究动作的依赖性、意图以及行动者的信念。问题被分为描述性、预测性、解释性和反事实性类型，以分析对空间、时间和因果关系的理解。该数据集的任务是视频问答。

This dataset, named EgoTaskQA, serves as a benchmark for human task understanding. By conducting question-answering tasks on real-world first-person perspective videos, it aims to investigate action dependencies, intentions, and agents' beliefs. The questions are categorized into four types: descriptive, predictive, explanatory, and counterfactual, to analyze the comprehension of spatial, temporal, and causal relationships. The core task of this dataset is video question answering.

搜集汇总

背景与挑战

背景概述

EgoTaskQA是一个用于评估模型在自我中心视频中理解人类任务能力的基准数据集。它基于增强的LEMMA数据集，标注了物体状态、关系和智能体信念，生成针对动作依赖、意图和智能体信念的多样化问题。数据集包含描述性、预测性、解释性和反事实查询，旨在诊断模型在空间、时间和因果方面的理解能力。