Name That Part: 3D Part Segmentation and Naming
Soumava Paul, Prakhar Kaushik, Ankit Vaidya, Anand Bhattad, Alan Yuille
2025-12-23
Summary
This paper focuses on breaking down 3D objects into their component parts and giving those parts meaningful names, like identifying the 'seat' of a chair or the 'wheel' of a car.
What's the problem?
Currently, it's hard to train computers to do this well because different datasets use different names and definitions for the same parts. Existing methods either don't label the parts at all, or can only find individual parts without understanding the whole object's structure. It's like trying to build with LEGOs when some instructions call a brick a 'stud' and others call it a 'block'.
What's the solution?
The researchers developed a method called ALIGN-Parts that treats part naming as a matching problem. It breaks objects into small 3D pieces called 'partlets' and then tries to match them to descriptions of parts using a combination of shape, how the part looks from different angles, and what the part is *for* (like 'supports weight' for a chair seat). They use a special 'text-alignment loss' to make sure the partlets and their descriptions exist in the same 'understanding space', allowing the system to potentially identify parts it hasn't specifically been trained on.
Why it matters?
This work is important because it creates a more consistent way to understand and label 3D parts, even across different datasets. It can be used to automatically annotate 3D models, which is a huge time saver, and it allows for identifying parts based on descriptions, even if the system hasn't seen that specific part before. They even created a unified list of over 1,700 different 3D parts by combining information from multiple existing datasets.
Abstract
We address semantic 3D part segmentation: decomposing objects into parts with meaningful names. While datasets exist with part annotations, their definitions are inconsistent across datasets, limiting robust training. Previous methods produce unlabeled decompositions or retrieve single parts without complete shape annotations. We propose ALIGN-Parts, which formulates part naming as a direct set alignment task. Our method decomposes shapes into partlets - implicit 3D part representations - matched to part descriptions via bipartite assignment. We combine geometric cues from 3D part fields, appearance from multi-view vision features, and semantic knowledge from language-model-generated affordance descriptions. Text-alignment loss ensures partlets share embedding space with text, enabling a theoretically open-vocabulary matching setup, given sufficient data. Our efficient and novel, one-shot, 3D part segmentation and naming method finds applications in several downstream tasks, including serving as a scalable annotation engine. As our model supports zero-shot matching to arbitrary descriptions and confidence-calibrated predictions for known categories, with human verification, we create a unified ontology that aligns PartNet, 3DCoMPaT++, and Find3D, consisting of 1,794 unique 3D parts. We also show examples from our newly created Tex-Parts dataset. We also introduce 2 novel metrics appropriate for the named 3D part segmentation task.