Research map / Detection and segmentation / Open-set and few-shot detection
ShapeMatcher: Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation
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 ShapeMatcher, a self-supervised framework that jointly learns shape canonicalization, part segmentation, region-weighted retrieval, and part-center-guided cage deformation for partial objects in arbitrary poses.
Problem
Retrieval-and-deformation (R&D) methods for 3D shape reconstruction assume pre-canonicalized targets and struggle with partial observations, so real-world scans in arbitrary poses need separate pose estimation.
Approach
Jointly train four associated modules (canonicalization via affine-invariant VN features, unsupervised part segmentation with part centers, region-weighted retrieval tokens, and part-center guided neural cage deformation) with partial-full and cross-task consistency losses so they supervise each other without annotations.
Key results
- On PartNet full shapes, ShapeMatcher averages 0.200 Chamfer Distance (x10^-2) versus 0.725 for U-RED + PE, the best baseline with pose estimation.
- On partial PartNet shapes at 25% occlusion, ShapeMatcher achieves 0.776 versus 1.625 for U-RED + PE.
- On real-world Scan2CAD (zero-shot), ShapeMatcher averages 0.375 versus 3.354 for U-RED + PE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ShapeMatcher instead of the strongest reported baseline | PartNet full-shape retrieval and deformation · PartNet full shapes | Chamfer Distance: 0.725 -> 0.200 | Helped |
| Use ShapeMatcher instead of the strongest reported baseline | ComplementMe retrieval and deformation · ComplementMe full shapes | Chamfer Distance: 1.999 -> 0.294 | Helped |
| Use ShapeMatcher instead of the strongest reported baseline | PartNet retrieval and deformation · PartNet 10% occlusion | Chamfer Distance: 1.147 -> 0.622 | Helped |
| Use ShapeMatcher instead of the strongest reported baseline | PartNet retrieval and deformation · PartNet 25% occlusion | Chamfer Distance: 1.625 -> 0.776 | Helped |
| Use ShapeMatcher instead of the strongest reported baseline | PartNet retrieval and deformation · PartNet 50% occlusion | Chamfer Distance: 3.628 -> 1.194 | Helped |
| Use ShapeMatcher instead of the strongest reported baseline | Zero-shot retrieval and deformation · Scan2CAD | Chamfer Distance: 3.354 -> 0.375 | Helped |
| Add translation decoupling (ablation) | ShapeMatcher canonicalization · PartNet | Chamfer Distance: 0.590 -> 0.506 | Helped |
| Add rotation decoupling (ablation) | ShapeMatcher canonicalization with translation decoupling · PartNet | Chamfer Distance: 0.506 -> 0.244 | Helped |
| Add scale decoupling (ablation) | ShapeMatcher canonicalization with translation and rotation decoupling · PartNet | Chamfer Distance: 0.244 -> 0.200 | Helped |
| Replace global-feature deformation with part-center-guided deformation (ablation) | ShapeMatcher with global retrieval · PartNet 25% occlusion | Chamfer Distance: 1.570 -> 1.431 | Helped |
| Replace global-feature retrieval with region-weighted retrieval (ablation) | ShapeMatcher with part-center-guided deformation · PartNet 25% occlusion | Chamfer Distance: 1.431 -> 0.776 | Helped |
| Replace global-feature deformation with part-center-guided deformation (ablation) | ShapeMatcher with region-weighted retrieval · PartNet 25% occlusion | Chamfer Distance: 0.973 -> 0.776 | Helped |
Limitations
- Limitations are discussed only in the supplementary material, which is not included in the provided text.
- Evaluated on a limited set of three categories (tables, chairs, cabinets).
Experimental setup
Datasets: PartNet ComplementMe Scan2CAD ShapeNet
Benchmarks: PartNet ComplementMe Scan2CAD
Metrics: Chamfer Distance
Baselines: Uy et al. U-RED Uy et al. + PE U-RED + PE
Builds on: Vector Neurons VN-MLP pose canonicalization Uy et al. R&D U-RED neural cage deformation
Closest papers by meaning
- ConDor: Self-Supervised Canonicalization of 3D Pose for Partial Shapes (CVPR 2022 · similarity 0.87)
- Joint Learning of 3D Shape Retrieval and Deformation (CVPR 2021 · similarity 0.87)
- KP-RED: Exploiting Semantic Keypoints for Joint 3D Shape Retrieval and Deformation (CVPR 2024 · similarity 0.87)
- One-shot 3D Object Canonicalization based on Geometric and Semantic Consistency (CVPR 2025 · similarity 0.86)
- Learning SO(3)-Invariant Semantic Correspondence via Local Shape Transform (CVPR 2024 · similarity 0.85)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.85)
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.