EQA v1

EmbodiedQA 数据集由 Georgia Institute of Technology 和 Facebook AI Research 联合开发，旨在推动智能体在三维环境中通过主动感知和导航回答视觉问题的研究。该数据集基于 House3D 环境构建，包含超过 5000 个问题，覆盖 750 个虚拟环境，涉及 45 种独特物体和 7 种房间类型。问题涵盖场景识别、颜色识别、空间关系推理等多种类型，以功能性程序形式生成，确保问题的多样性和答案分布的平衡。数据集通过程序化方法生成问题与答案，避免数据偏差，同时结合人类标注验证环境布局的合理性。EmbodiedQA 数据集主要用于训练和评估智能体的导航、视觉感知和语言理解能力，推动多模态交互智能体的发展。

EmbodiedQA is a dataset co-developed by the Georgia Institute of Technology and Facebook AI Research, aimed at advancing research on intelligent agents that answer visual questions via active perception and navigation within 3D environments. Constructed based on the House3D environment, this dataset contains more than 5,000 questions spanning 750 virtual environments, covering 45 distinct object types and 7 room categories. The questions encompass diverse tasks including scene recognition, color recognition, spatial relationship reasoning and others, and are generated through functional programming to ensure both the diversity of questions and the balanced distribution of answers. The dataset adopts procedural generation for questions and answers to mitigate data bias, and combines human annotations to validate the rationality of environment layouts. The EmbodiedQA dataset is primarily utilized for training and evaluating the navigation, visual perception and language comprehension abilities of intelligent agents, thereby promoting the development of multimodal interactive AI agents.

搜集汇总

背景与挑战

背景概述

EQA v1.0是一个包含9,000个问题的数据集，覆盖774个环境，主要涉及位置、颜色和地点介词问题，并提供目标对象和房间的边界框标注。数据集还包含人类在AMT上收集的导航轨迹和答案，涉及1,162个问题和70个环境。