Struct2D-Set
收藏资源简介:
Struct2D-Set是一个大规模的指令微调数据集,包含来自3D室内场景的20万个细粒度的QA对,涵盖八个与具身AI相关的空间推理类别。该数据集通过自动化流程自动生成,利用原始3D数据集中提供的真实对象标注。Struct2D-Set旨在帮助大型多模态模型(LMMs)获得丰富的3D空间推理技能,通过指令微调仅使用结构化2D表示,无需直接访问3D点云。数据集的构建包括使用ChatGPT丰富QA对并提供推理线索,以及人类参与的审核流程来进一步验证数据集。该数据集的发布旨在支持未来的研究,并推动LMMs在空间推理任务中的发展。
Struct2D-Set is a large-scale instruction tuning dataset containing 200,000 fine-grained QA pairs from 3D indoor scenes, covering eight embodied AI-related spatial reasoning categories. This dataset is automatically generated through an automated pipeline, leveraging real object annotations provided in the original 3D dataset. Struct2D-Set aims to equip large multimodal models (LMMs) with rich 3D spatial reasoning skills via instruction tuning using only structured 2D representations, without direct access to 3D point clouds. The construction of the dataset involves enriching QA pairs with ChatGPT to provide reasoning clues, as well as a human-in-the-loop review process to further validate the dataset. The release of this dataset is intended to support future research and advance the development of LMMs in spatial reasoning tasks.
Struct2D数据集概述
数据集基本信息
- 数据集名称:Struct2D
- 代码发布:用于论文《Struct2D: A Perception-Guided Framework for Spatial Reasoning in Large Multimodal Models》
数据集用途
- 专注于空间推理任务
- 面向大型多模态模型(LMMs)开发
技术特点
- 采用感知引导框架
- 针对二维空间结构推理问题设计

- 1Struct2D: A Perception-Guided Framework for Spatial Reasoning in Large Multimodal Models东北大学,微软研究院,南加州大学,加州大学圣克鲁兹分校 · 2025年



