SpatialCrafter 数据集 (未命名,指代该大规模场景数据集)
收藏资源简介:
SpatialCrafter 论文构建了首个专为图像到场景生成任务设计的大规模混合数据集,包含约115K个高保真3D场景。该数据集整合了合成室内场景渲染与真实世界视频,通过深度估计和相机位姿重建生成一致的3D几何,并经过严格筛选以剔除不可靠样本。数据覆盖室内外多样化环境,提供精确的几何标注,旨在解决单图像场景生成中缺乏合适训练数据的瓶颈,从而支持3D一致的世界建模,应用于游戏、机器人、虚拟现实等领域。
The paper *SpatialCrafter* presents the first large-scale hybrid dataset specifically designed for image-to-scene generation tasks, containing approximately 115K high-fidelity 3D scenes. This dataset integrates synthetic indoor scene renderings and real-world videos, generates consistent 3D geometry via depth estimation and camera pose reconstruction, and undergoes rigorous filtering to eliminate unreliable samples. It covers diverse indoor and outdoor environments and provides precise geometric annotations. The dataset aims to address the bottleneck of lacking suitable training data for single-image scene generation, thereby supporting 3D-consistent world modeling, with applications in fields such as gaming, robotics, virtual reality, and more.





