基于语义标注的真实3D空间数据集
收藏资源简介:
本文提出了一个名为“基于相关信息的期望计算”的模型(CECI),用于在信念场景图上估计常识场景组成。该数据集基于语义标注的真实3D空间数据集,通过筛选和分组数据集中的1659个语义类别,生成35个精选的语义类别标签。数据集包括真实场景的3D空间图,每个节点具有语义类别标签、真实位置和真实尺寸等属性。此外,论文还介绍了使用大型语言模型(LLM)开发空间本体,以增强模型的学习能力。
This paper proposes a model named Correlation Information-based Expectation Calculation (CECI) for estimating commonsense scene composition on belief scene graphs. This dataset is built upon a real-world 3D spatial dataset with semantic annotations. By screening and grouping the 1659 semantic categories in the original dataset, 35 curated semantic category labels are generated. The dataset contains 3D spatial graphs of real-world scenes, where each node is equipped with attributes including semantic category labels, real-world positions and real-world dimensions. Additionally, this paper also presents the development of spatial ontologies using Large Language Models (LLMs) to improve the model's learning ability.

- 1Estimating Commonsense Scene Composition on Belief Scene Graphs瑞典吕勒奥工业大学计算机、电气和空间工程学院机器人与人工智能团队 · 2025年



