Research map / Point clouds / Shape and surface completion
Point Cloud Part Editing: Segmentation, Generation, Assembly, and Selection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SGAS, a four-stage (segmentation, generation, assembly, selection) point cloud part editing model using per-part GANs on pre-trained part autoencoders with a global discriminator and part masking/selection, plus a new TMDS diversity metric.
Problem
Point cloud part editing needs diverse edited parts, fidelity to unedited parts, and high-quality assembled results; prior methods entangle parts and lack inter-part constraints, requiring manual selection.
Approach
Disentangle parts by independently fitting part-level feature distributions with per-part GANs over pre-trained part autoencoders, conditioned on unedited parts via AdaIN, and enforce assembly constraints with a global discriminator on all part features plus a learned part-selection threshold.
Key results
- Diversity part editing on Chair/Lamp/Table: SGAS average MMD 1.55 vs MSC-cGAN 2.14 and average TMD 5.63 vs 4.84.
- Unsupervised generation on Chair: SGAS COV-EMD 58 vs best baseline 32 (EditVAE M=7), with EMD-MMD 0.076 best among compared models.
- Ablation on number of parts: n=6 gives best MMD-EMD (0.039) and COV-EMD (58) on Airplane.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SGAS instead of MSC-cGAN | Point cloud part editing · PartNet.v0.Merged Chair | MMD: 1.62 -> 1.33 | Helped |
| Use SGAS instead of MSC-cGAN | Point cloud part editing · PartNet.v0.Merged Average | TMD: 4.84 -> 5.63 | Mixed |
| Set α:β to higher diversity weighting (1:1 to 10:1) (ablation) | SGAS part editing · PartNet.v0.Merged | TMD: 4.36 -> 7.68 | Helped |
| Set α:β to higher diversity weighting (1:1 to 10:1) (ablation) | SGAS part editing · PartNet.v0.Merged | MMD: 1.33 -> 1.40 | Hurt |
| Use 6 parts rather than 2 (ablation) | SGAS generation · ShapeNet-Partseg Airplane | COV EMD: 47 -> 58 | Helped |
| Use 6 parts rather than 13 (ablation) | SGAS generation · ShapeNet-Partseg Airplane | COV EMD: 53 -> 58 | Helped |
| Vary the number of parts from 2 to 13 (ablation) | SGAS generation · ShapeNet-Partseg Airplane | JSD: 0.042; 0.039; 0.040; 0.036; 0.040; 0.035 | Mixed |
Surprising findings
- Part selection is training-free: parts whose latent features fall near zero are simply not decoded, letting the model omit parts such as chair arms.
Limitations
- Without part offset networks, relies on fixed spatial positions of parts, so works only for objects with consistent prototypes (e.g., cannot mix ceiling and table lamps).
- Performance limited by pre-trained part autoencoders; embeddings work better when parts are normalized.
- Unsupervised generation can show gaps between parts, requiring point-count expansion heuristic.
Experimental setup
Datasets: PartNet PartNet.v0.Merged ShapeNet-Partseg ScanNet
Benchmarks: PartNet.v0.Merged ShapeNet-Partseg ScanNet
Metrics: MMD TMD TMDS JSD COV UHD
Baselines: MSC-cGAN TreeGAN MRGAN EditVAE SP-GAN
Builds on: l-GAN Wasserstein GAN with gradient penalty AdaIN EditVAE
Closest papers by meaning
- SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation (CVPR 2025 · similarity 0.85)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.85)
- Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.84)
- Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing (CVPR 2024 · similarity 0.84)
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation (ICML 2025 · similarity 0.84)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.