PartNeXt: A Next-Generation Dataset for Fine-Grained and Hierarchical 3D Part Understanding
Penghao Wang, Yiyang He, Xin Lv, Yukai Zhou, Lan Xu, Jingyi Yu, Jiayuan Gu
2025-10-29
Summary
This paper introduces PartNeXt, a new and improved dataset for teaching computers to understand 3D objects by recognizing their individual parts.
What's the problem?
Current datasets used for this kind of research, like PartNet, have limitations. They often use simple, uncolored 3D models and require experts to carefully label every part, making it hard to create large, useful datasets. Existing methods struggle to accurately identify very specific parts within objects, and there wasn't a good way to test if AI models could actually *understand* what parts were being asked about in a question.
What's the solution?
The researchers created PartNeXt, a dataset with over 23,000 detailed, colorful 3D models. These models are labeled with the names of their parts in a structured way, showing how parts relate to each other. They then tested existing AI models on this dataset using two tasks: identifying parts in general, and answering questions about specific parts of objects. They also showed that training a particular AI model, Point-SAM, with PartNeXt made it perform much better than when trained with the older PartNet dataset.
Why it matters?
PartNeXt is important because it provides a better resource for developing AI that can understand the world like humans do – by recognizing objects not just as whole things, but as collections of parts. The dataset’s size, detail, and the new question-answering task will help push the field of computer vision and robotics forward, allowing for more advanced and capable AI systems.
Abstract
Understanding objects at the level of their constituent parts is fundamental to advancing computer vision, graphics, and robotics. While datasets like PartNet have driven progress in 3D part understanding, their reliance on untextured geometries and expert-dependent annotation limits scalability and usability. We introduce PartNeXt, a next-generation dataset addressing these gaps with over 23,000 high-quality, textured 3D models annotated with fine-grained, hierarchical part labels across 50 categories. We benchmark PartNeXt on two tasks: (1) class-agnostic part segmentation, where state-of-the-art methods (e.g., PartField, SAMPart3D) struggle with fine-grained and leaf-level parts, and (2) 3D part-centric question answering, a new benchmark for 3D-LLMs that reveals significant gaps in open-vocabulary part grounding. Additionally, training Point-SAM on PartNeXt yields substantial gains over PartNet, underscoring the dataset's superior quality and diversity. By combining scalable annotation, texture-aware labels, and multi-task evaluation, PartNeXt opens new avenues for research in structured 3D understanding.