Voxel51/partscan
收藏资源简介:
PartScan是一个用于细粒度3D部件分割的数据集,对具身AI系统至关重要,这些系统需要与物体的特定功能部件(例如抽屉把手而非整个柜子)进行交互。由于获取密集的部件级3D注释是一个主要瓶颈,PinPoint3D引入了一个3D数据合成流程,生成了一个大规模的场景级数据集,该数据集在稀疏的、真实世界风格的扫描上具有密集的部件注释。PartScan已被解析为FiftyOne的3D点云数据集,包含场景级扫描和密集的每点部件级注释。它是为PinPoint3D框架引入的合成数据集,该框架用于通过少量用户点击实现细粒度、多粒度的3D部件分割。每个样本是一个彩色点云场景片段,在FiftyOne App的3D查看器中渲染。
Fine-grained 3D part segmentation is crucial for embodied AI systems that must interact with specific functional components of an object (e.g. a drawer handle rather than the whole cabinet). Acquiring dense, part-level 3D annotations is a major bottleneck, so PinPoint3D introduces a 3D data-synthesis pipeline that produces a large-scale, scene-level dataset with dense part annotations on sparse, real-world-style scans. PartScan has been parsed as a FiftyOne 3D point cloud dataset of scene-level scans with dense, per-point part-level annotations. It is the synthesized dataset introduced for PinPoint3D, a framework for fine-grained, multi-granularity 3D part segmentation from a few user clicks. Each sample is a colored point cloud of one scene fragment, rendered in the FiftyOne Apps 3D viewer.




