TO-Scene
收藏资源简介:
TO-Scene是由香港中文大学(深圳)创建的一个大规模数据集,专注于理解3D桌面场景。该数据集包含20,740个场景,通过设计一个高效且可扩展的框架来收集数据,其中开发了一个众包用户界面,用于将来自ModelNet和ShapeNet的CAD对象转移到ScanNet中的真实桌面上。数据集的创建过程涉及模拟合成对象到真实扫描,并自动进行标注。TO-Scene数据集包含三个变体,用于不同场景的测试,并且提供了一个真实的扫描测试集TO-Real,以验证数据集的实际应用价值。该数据集主要用于解决室内场景解析中的问题,特别是在处理小型桌面实例时的挑战。
TO-Scene is a large-scale dataset focused on 3D desktop scene understanding, created by The Chinese University of Hong Kong, Shenzhen. Comprising 20,740 scenes, the dataset was collected via an efficient and scalable framework, where a crowdsourcing user interface was developed to transfer CAD objects sourced from ModelNet and ShapeNet onto real desktop setups in ScanNet. The dataset creation process involves simulating synthetic objects onto real scans and performing automatic annotation. The TO-Scene dataset includes three variants for testing across different scenarios, and a real scan test set named TO-Real is provided to validate its practical application value. This dataset is primarily intended to address challenges in indoor scene parsing, especially those arising when handling small desktop instances.




