3D part amodal segmentation--decomposing a 3D shape into complete, semantically meaningful parts, even when occluded--is a challenging but crucial task for 3D content creation and understanding. Existing 3D part segmentation methods only identify visible surface patches, limiting their utility. Inspired by 2D amodal segmentation, we introduce this novel task to the 3D domain and propose a practical, two-stage approach, addressing the key challenges of inferring occluded 3D geometry, maintaining global shape consistency, and handling diverse shapes with limited training data. First, we leverage existing 3D part segmentation to obtain initial, incomplete part segments. Second, we introduce PartComp, a novel diffusion-based model, to complete these segments into full 3D parts. PartComp utilizes a specialized architecture with local attention to capture fine-grained part geometry and global shape context attention to ensure overall shape consistency. We introduce new benchmarks based on the ABO and PartObjaverse-Tiny datasets and demonstrate that PartComp significantly outperforms state-of-the-art shape completion methods. Incorporating PartComp with existing segmentation techniques, we achieve promising results on 3D part amodal segmentation, opening new avenues for applications in geometry editing, animation, and material assignment.
By representing a part with the same number of tokens as the overall object, we can achieve geometry super-resolution.
Given a whole 3D shape and a corresponding surface segmentation mask,
PartComp encodes these inputs into latent tokens, using context-aware attention to
capture global shape context and local attention to capture local part detailed features
and position mapping. These tokens are used as conditions and injected into the part
diffusion model via cross-attention respectively. During training, noise is added to
complete 3D parts, and the model learns to denoise them and recover the original complete part.
@article{yang2025holopart,
title={HoloPart: Generative 3D Part Amodal Segmentation},
author={Yang, Yunhan and Guo, Yuan-Chen and Huang, Yukun and Zou, Zi-Xin and Yu, Zhipeng and Li, Yangguang and Cao, Yan-Pei and Liu, Xihui},
journal={arXiv preprint arXiv:2504.07943},
year={2025}
}