ScanQA
收藏资源简介：
ScanQA 数据集由京都大学、ATR、RIKEN AIP 等机构的研究人员共同创建，旨在推动 3D 空间理解任务的发展。该数据集包含超过 41,000 个问答对，覆盖 800 个室内场景，基于 ScanNet 数据集构建。问答对通过自动化生成和人工编辑相结合的方式生成，涵盖了多种类型的自然语言问题，如物体识别、颜色、位置等。数据集的创建过程包括自动从 ScanRefer 数据集的描述中生成问题、人工筛选和编辑问题以及收集自由形式的答案。ScanQA 数据集的应用领域广泛，主要用于训练和评估能够理解 3D 空间信息并回答相关问题的模型。它解决了传统 2D 图像问答模型在 3D 空间理解上的局限性，如物体对齐、方向理解、遮挡问题以及多图像中的物体识别问题。该数据集为开发能够在 3D 环境中进行问答的智能模型提供了丰富的资源，有助于推动机器人导航、虚拟现实交互等领域的研究。
The ScanQA dataset was co-created by researchers from institutions including Kyoto University, ATR, and RIKEN AIP, aiming to advance the development of 3D spatial understanding tasks. This dataset contains over 41,000 question-answer pairs covering 800 indoor scenes, and is constructed based on the ScanNet dataset. The question-answer pairs are generated via a hybrid workflow of automated generation and manual editing, covering various types of natural language questions such as object recognition, color inquiry, spatial location, and more. The dataset development pipeline includes three key steps: automatically generating questions from the descriptions of the ScanRefer dataset, manually screening and revising these generated questions, and collecting free-form answers. The ScanQA dataset has broad application scenarios, and is mainly used to train and evaluate models that can understand 3D spatial information and answer related questions. It addresses the limitations of traditional 2D visual question answering models in 3D spatial understanding tasks, including object alignment, orientation comprehension, occlusion issues, and object recognition across multiple images. This dataset provides a rich resource for developing intelligent models capable of conducting question answering in 3D environments, and promotes research in fields such as robotic navigation and virtual reality interaction.
ScanQA 数据集概述
数据集简介
ScanQA 是一个用于3D空间场景理解的问答数据集，旨在通过3D场景中的视觉信息回答文本问题。该数据集由41,000多个问答对组成，涵盖800个室内场景，这些场景来自ScanNet数据集。
数据集特点
- 任务类型：3D问答（3D-QA），模型需要从3D场景的RGB-D扫描中获取视觉信息，并回答关于该场景的文本问题。
- 问题类型：自由形式的问答对，问题与3D场景中的对象相关联。
- 数据规模：超过41,000个问答对，来自800个室内场景。
数据集使用
- 训练：使用RGB值训练ScanQA模型。
- 推理：对训练好的模型进行评估、评分和预测。
- 基准测试：ScanQA基准测试托管在EvalAI平台上，用户可以通过提交预测结果进行评估。
数据集引用
如果使用该数据集进行研究，请引用以下论文： bibtex @inproceedings{azuma_2022_CVPR, title={ScanQA: 3D Question Answering for Spatial Scene Understanding}, author={Azuma, Daichi and Miyanishi, Taiki and Kurita, Shuhei and Kawanabe, Motoaki}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year={2022} }
许可证
ScanQA数据集采用Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License许可。
致谢
感谢以下项目的贡献：
- facebookresearch/votenet 提供了3D对象检测代码。
- daveredrum/ScanRefer 提供了3D定位代码库。
