IL3D
收藏资源简介:
IL3D是一个大规模室内布局数据集,专门为大型语言模型驱动的3D场景生成而设计,旨在解决室内布局设计中高质量训练数据的迫切需求。该数据集包含27,816个室内布局,涵盖18种常见的房间类型,以及29,215个高保真3D物体资产库。IL3D还富含实例级别的自然语言注释,以支持视觉-语言任务的鲁棒多模态学习。数据集的构建基于USD格式,使用USDZ格式的3D物体资产和USDA格式的房间布局。这种格式的显著特点是文本可读性,即大型语言模型可以直接从3D场景模型中读取场景中物体的信息。此外,IL3D提供了实例级别的自然语言描述,并支持多种数据格式,包括语义点云、3D边界框、多视图RGB图像、深度图、法线图和语义掩码,确保与各种下游视觉任务的兼容性。
IL3D is a large-scale indoor layout dataset specifically developed for large language model (LLM)-enabled 3D scene generation, aiming to address the critical demand for high-quality training data in indoor layout design. This dataset comprises 27,816 indoor layouts spanning 18 common room categories, alongside a library of 29,215 high-fidelity 3D object assets. IL3D also features rich instance-level natural language annotations to support robust multimodal learning for vision-language tasks. The dataset is constructed based on the USD format, using USDZ-formatted 3D object assets and USDA-formatted room layouts. A prominent characteristic of this format is its text readability, allowing large language models to directly read object information within the 3D scene from the scene model. Furthermore, IL3D provides instance-level natural language descriptions and supports multiple data formats including semantic point clouds, 3D bounding boxes, multi-view RGB images, depth maps, normal maps and semantic masks, ensuring compatibility with various downstream visual tasks.



