SQA3D
收藏资源简介:
SQA3D是由北京通用人工智能研究院、加州大学洛杉矶分校、清华大学和北京大学联合发布的数据集,旨在评估智能体对3D场景的理解能力。该数据集包含650个ScanNet场景中的6.8k种独特情境,以及20.4k个情境描述和33.4k个多样化的问题,涵盖空间关系、常识推理和多跳推理等多种能力。数据集通过众包方式从亚马逊MTurk收集情境描述和问题,并经过严格的筛选和平衡处理。SQA3D的创建过程分为情境识别、问题准备和答案收集三个阶段,以确保数据质量和多样性。该数据集可用于研究智能体在复杂3D环境中的情境理解与推理能力,推动具身人工智能的发展。
SQA3D is a dataset jointly released by the Beijing Academy of General Intelligence, University of California, Los Angeles, Tsinghua University, and Peking University, aiming to evaluate the 3D scene understanding capabilities of AI Agents. This dataset contains 6.8k unique scenarios across 650 ScanNet scenes, alongside 20.4k scenario descriptions and 33.4k diverse questions covering multiple abilities including spatial relations, commonsense reasoning, and multi-hop reasoning. The scenario descriptions and questions were collected via crowdsourcing on Amazon MTurk, and underwent rigorous filtering and balancing procedures. The creation pipeline of SQA3D is divided into three stages: scenario identification, question preparation, and answer collection, to guarantee data quality and diversity. This dataset can be employed to study the scenario understanding and reasoning capabilities of AI Agents in complex 3D environments, thus promoting the development of embodied artificial intelligence.




