OpenNovelty.org

HoloPart: Generative 3D Part Amodal Segmentation

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
3D Generation3D Segmentation3D Part
Abstract:

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 HoloPart, a novel diffusion-based model, to complete these segments into full 3D parts. HoloPart 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 HoloPart significantly outperforms state-of-the-art shape completion methods. By incorporating HoloPart 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.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces 3D part amodal segmentation—decomposing 3D shapes into complete, semantically meaningful parts even under occlusion—and proposes HoloPart, a diffusion-based model for part completion. Within the taxonomy, this work resides in the 'Diffusion-Based Amodal Reconstruction' leaf, which contains five papers total. This leaf sits under 'Amodal Completion and Shape Reconstruction Methods,' indicating a moderately active research direction focused on generative approaches to occlusion reasoning. The taxonomy reveals that diffusion-based methods represent one of several competing paradigms for amodal reconstruction, alongside VAE/GAN-based approaches and multi-view techniques.

The taxonomy structure shows neighboring leaves addressing related but distinct challenges: 'Other Generative Amodal Approaches' explores non-diffusion generative models, while 'Single-View Amodal Reconstruction' focuses on RGB-based inference without the part-level decomposition emphasis. The 'Part Segmentation with Amodal Reasoning' leaf under a separate branch addresses similar occlusion handling but from a segmentation-first perspective rather than generative completion. The scope notes clarify that this work differs from 'Part Segmentation without Amodal Reasoning' by explicitly modeling occluded geometry, and from 'Part-Aware 3D Generation' by operating on existing shapes rather than generating from scratch.

Among 30 candidate papers examined, the contribution-level analysis reveals mixed novelty signals. The task formulation and benchmarks (Contribution 1) show no clear refutation across 10 candidates, suggesting this specific problem framing may be relatively underexplored. The HoloPart diffusion architecture (Contribution 2) similarly shows no refutation among 10 examined papers, indicating potential architectural novelty within the diffusion-based reconstruction space. However, the two-stage approach combining segmentation and completion (Contribution 3) found one refutable candidate among 10 examined, suggesting some prior work explores similar decomposition-then-completion pipelines, though the scale of overlap remains limited given the search scope.

Based on this limited analysis of 30 semantically similar papers, the work appears to occupy a moderately explored niche within diffusion-based 3D reconstruction. The task-level novelty seems stronger than the methodological approach, with the two-stage pipeline showing some precedent. The taxonomy context suggests this sits at the intersection of two active areas—diffusion-based generation and part-level reasoning—where the specific combination may offer incremental contributions. A more exhaustive search beyond the top-30 semantic matches woul

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: 3D part amodal segmentation with complete geometry inference. This field addresses the challenge of segmenting object parts in 3D while inferring their complete geometry even when portions are occluded or missing. The taxonomy reveals a rich landscape organized around several complementary perspectives. Amodal Completion and Shape Reconstruction Methods focus on recovering full object shapes from partial observations, often leveraging diffusion models or neural implicit representations to hallucinate occluded regions. Part-Level Segmentation and Decomposition emphasizes breaking objects into semantic components, with works like PartSAM[4] and Omnipart[6] exploring how to identify and delineate individual parts. Meanwhile, 2D Amodal Segmentation and Image-Based Methods tackle occlusion reasoning in image space, and Instance Segmentation and Tracking with Occlusion Handling extends these ideas to temporal scenarios. Geometry-Guided Segmentation and Reconstruction exploits geometric priors to constrain and improve predictions, while Domain-Specific Applications demonstrate how these techniques adapt to specialized contexts such as agriculture or medical imaging. Recent activity has concentrated on diffusion-based approaches for amodal reconstruction, where generative models learn to complete missing geometry in a probabilistically principled manner. HoloPart[0] sits within this Diffusion-Based Amodal Reconstruction cluster, alongside works like pix2gestalt[1], Amodal3r[8], and AmodalGen3D[10], all of which harness diffusion priors to infer occluded part structures. Compared to Sequential Amodal[27], which may adopt iterative refinement strategies, HoloPart[0] emphasizes end-to-end generation of complete part geometries. The interplay between part-level decomposition methods such as Stable Part Diffusion[5] and holistic shape completion remains an open question: whether to first segment then complete, or to jointly infer parts and their full extents. Domain-specific challenges, illustrated by works on occluded fruits or tree branches, highlight the need for robust occlusion handling across varied real-world scenarios, situating HoloPart[0] as part of a broader effort to unify geometric reasoning with modern generative modeling.

Claimed Contributions

3D part amodal segmentation task and benchmarks

The authors define a new task called 3D part amodal segmentation that decomposes 3D shapes into complete semantic parts rather than visible surface patches. They establish two evaluation benchmarks using the ABO and PartObjaverse-Tiny datasets to enable future research in this area.

10 retrieved papers
HoloPart diffusion-based model for 3D part shape completion

The authors introduce HoloPart, a diffusion-based model designed specifically for completing 3D part shapes. The model incorporates local attention to capture fine-grained part geometry and context-aware attention to maintain global shape consistency, while leveraging pretrained 3D generative priors to handle limited training data.

10 retrieved papers
Two-stage approach combining segmentation and part completion

The authors propose a practical two-stage pipeline where existing 3D part segmentation methods provide initial incomplete surface patches, followed by their HoloPart model that completes these segments into full 3D parts with proper geometry and semantic consistency.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

3D part amodal segmentation task and benchmarks

The authors define a new task called 3D part amodal segmentation that decomposes 3D shapes into complete semantic parts rather than visible surface patches. They establish two evaluation benchmarks using the ABO and PartObjaverse-Tiny datasets to enable future research in this area.

Contribution

HoloPart diffusion-based model for 3D part shape completion

The authors introduce HoloPart, a diffusion-based model designed specifically for completing 3D part shapes. The model incorporates local attention to capture fine-grained part geometry and context-aware attention to maintain global shape consistency, while leveraging pretrained 3D generative priors to handle limited training data.

Contribution

Two-stage approach combining segmentation and part completion

The authors propose a practical two-stage pipeline where existing 3D part segmentation methods provide initial incomplete surface patches, followed by their HoloPart model that completes these segments into full 3D parts with proper geometry and semantic consistency.

Table of Contents