Abstract
We propose PartField, a feedforward approach for learning part-based 3Dfeatures, which captures the general concept of parts and their hierarchywithout relying on predefined templates or text-based names, and can be appliedto open-world 3D shapes across various modalities. PartField requires only a 3Dfeedforward pass at inference time, significantly improving runtime androbustness compared to prior approaches. Our model is trained by distilling 2Dand 3D part proposals from a mix of labeled datasets and image segmentations onlarge unsupervised datasets, via a contrastive learning formulation. Itproduces a continuous feature field which can be clustered to yield ahierarchical part decomposition. Comparisons show that PartField is up to 20%more accurate and often orders of magnitude faster than other recentclass-agnostic part-segmentation methods. Beyond single-shape partdecomposition, consistency in the learned field emerges across shapes, enablingtasks such as co-segmentation and correspondence, which we demonstrate inseveral applications of these general-purpose, hierarchical, and consistent 3Dfeature fields. Check our Webpage!https://research.nvidia.com/labs/toronto-ai/partfield-release/