Abstract
We study open-world part segmentation in 3D: segmenting any part in anyobject based on any text query. Prior methods are limited in object categoriesand part vocabularies. Recent advances in AI have demonstrated effectiveopen-world recognition capabilities in 2D. Inspired by this progress, wepropose an open-world, direct-prediction model for 3D part segmentation thatcan be applied zero-shot to any object. Our approach, called Find3D, trains ageneral-category point embedding model on large-scale 3D assets from theinternet without any human annotation. It combines a data engine, powered byfoundation models for annotating data, with a contrastive training method. Weachieve strong performance and generalization across multiple datasets, with upto a 3x improvement in mIoU over the next best method. Our model is 6x to over300x faster than existing baselines. To encourage research in general-categoryopen-world 3D part segmentation, we also release a benchmark for generalobjects and parts. Project website: https://ziqi-ma.github.io/find3dsite/