This repository contains the data used for the work "PartSDF: Part-Based Implicit Neural Representation for Composite 3D Shape Parametrization and Optimization", as well as the checkpoints of the pre-trained models.

The data consists in watertight meshes, their decomposition into parts, the poses of each parts, and the train/test splits used in the main work. They consists in processed meshes from ShapeNet (for cars and chairs), using the PartNet decomposition, and mixers from HybridSDF. Note that the 3D samples, with their accompanying SDF values and part labels, are not included because of their large size.

The checkpoints contain the experimental directory of the pre-trained models for each category present in the data and main text with: the network weights, the latent vectors and poses of the training shapes, and the specification files. These are best used with the official codebase of the work, see below for the Abstract and structure of the present dataset.

Abstract

Accurate 3D shape representation is essential in engineering applications such as design, optimization, and simulation. In practice, engineering workflows require structured, part-based representations, as objects are inherently designed as assemblies of distinct components. However, most existing methods either model shapes holistically or decompose them without predefined part structures, limiting their applicability in real-world design tasks. We propose PartSDF, a supervised implicit representation framework that explicitly models composite shapes with independent, controllable parts while maintaining shape consistency. Thanks to its simple but innovative architecture, PartSDF outperforms both supervised and unsupervised baselines in reconstruction and generation tasks. We further demonstrate its effectiveness as a structured shape prior for engineering applications, enabling precise control over individual components while preserving overall coherence. Code available at https://github.com/cvlab-epfl/PartSDF.

Data Structure

The data for each category follows this file structure:

dataset/
├── meshes/
├── parts/
│   ├── meshes/
│   ├── parameters/        ← Pose parameters of the parts
│   └── primitives/        ← Meshes of the parts primitives
└── splits/                ← Split files (each is a list of instance names)

Note that the 3D samples with SDF values and part labels are not included for memory constraints. Please refer to our code repository to generate them.

Checkpoint Structure

The checkpoint for each model follows this file structure:

dataset/
├── latent/
│   ├── latents_2000.pth     ← Training latents at epoch 2000
│   └── poses.pth            ← Training poses
├── log/
│   └── history.pth          ← Training history
├── model/│
└── model_2000.pth           ← Model weights at epoch 2000
└── specs.json               ← Specification file for the experiment