PartNet - A Large-scale Benchmark for Fine-grained and Hierarchical Part-level 3D Object Understanding - Supplementary Material

Sep 11, 2026

PartNet supplementary material expands the main paper with dataset visualizations and statistics, including fine-grained and hierarchical segmentation examples, shape annotation counts, semantic part statistics, and tree-depth measures. It further details expert-defined hierarchical template design, visualization, and the template refinement process used to reduce cross-annotator inconsistencies. The document also describes architectures and training settings for multiple shape segmentation tasks, supporting reproducible fine-grained and hierarchical 3D part understanding.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages14

Reading Time5 min

AccessFree to download

Overview

More Dataset Visualization and Statistics

More Fine-grained Segmentation Visualization

More Hierarchical Segmentation Visualization

Shape Statistics

Part Statistics

More Template Design Details and Visualization

Template Design Details

Template Refinement Details

PartNet: A Large-scale Benchmark for Fine-grained and Hierarchical Part-level 3D Object Understanding Supplementary Material Kaichun Mo 1 Shilin Zhu2 Angel X. Chang3 Li Yi 1 Subarna Tripathi4 Leonidas J. Guibas 1 Hao Su2 1 Stanford University 2University of California San Diego 3 Simon Fraser University 4Intel AI Lab [https://cs.stanford.edu/](https://cs.stanford.edu/)~kaichun/partnet/ A. Overview This document provides additional dataset visualization and statistics (Sec B), hierarchical template design details and visualization (Sec C), and the architectures and training details for the three shape segmentation tasks (Sec D), to the main paper. B. More Dataset Visualization and Statistics We present more visualization and statistics over the proposed PartNet dataset. B.1. More Fine-grained Segmentation Visualization Figure 5 and 6 show more visualization for ﬁne-grained instance-level segmentation annotations in PartNet. We observe the complexity of the annotated segmentation and the heterogeneous variation of shapes within each object category. B.2. More Hierarchical Segmentation Visualization Figure 7, 8 and 9 show more visualization for example hierarchical instance-level segmentation annotations in PartNet. We visualize the tree-structure of the hierarchical segmentation annotation with the 2D part renderings associated to the tree nodes. B.3. Shape Statistics We report the statistics for the number of annotations, unique shapes and shapes that we collect multiple human annotations in Figure 1. B.4. Part Statistics We report the statistics for the number of part semantics for each object category in Figure 2. We also present the statistics for the maximum and median number of part instances per shape for each object category in Figure 3. We Figure 1 . PartNet shape statistics. We report the statistics for the number of annotations, unique shapes and shapes that we collect multiple human annotations. report the statistics for the maximum and median tree depth for each object category in Figure 4. C. More Template Design Details and Visualization We provide more details and visualization for the expertdeﬁned hierarchical templates to guide the hierarchical segmentation annotation and the template reﬁnement procedure to resolve annotation inconsistencies. C.1. Template Design Details We design templates according to the rules of thumb that we describe in the main paper. We also consulted many online references 1 that describe object parts (often for manufacturing and assembly) and previous works that relate language to the shapes [2] as guides for the design of our 1 E.g. [http://www.props.eric-hart.com/resources/parts-of-a-chair/](http://www.props.eric-hart.com/resources/parts-of-a-chair/) . Figure 2 . PartNet part semantics statistics. We report the statistics for the number of part semantics for each object category. Figure 3 . PartNet part instance statistics. We report the statistics for the maximum and median number of part instances per shape for each object category. template. To ensure that our templates cover most of the shape variations and part semantics of each object category, we generated a t-SNE [5] visualization for the entire shape space to study the shape variation. We trained an autoencoder based on the shape geometry within each object category to obtain shape embeddings for the t-SNE visualization. Although we try to cover the most common part semantics in our templates, it is still not easy to cover all possible object parts. Thus, we allow annotators to deviate from the templates and deﬁne their own parts and segmentation structures. Among all the annotated part instances, 1.3% of them are deﬁned by the annotators. In the raw annotation, 13.1% of shapes contained user-deﬁned part labels. Our analysis shows that our template designs are able to cover most of the ShapeNet [1] shapes. Of the 27,260 shapes we collected in total, our annotators successf

FAQs

What additional dataset insights does the document provide beyond the main paper?

It adds dataset visualization and statistics, including fine-grained and hierarchical segmentation examples plus quantitative statistics such as annotations, unique shapes, part semantics, and tree depth.

How are hierarchical template design details and refinement addressed?

The document explains how hierarchical templates are defined and visualized, and how a refinement procedure and confusion analysis over multiple annotations help diagnose and reduce annotation inconsistencies.

What is the purpose of allowing annotators to deviate from predefined templates?

Annotators are permitted to define their own parts and segmentation structures when coverage is insufficient, helping capture additional object-part variations not fully represented by the templates.

Sep 11, 2026

Uploaded by supergirl