Research map / Point clouds / Shape and surface completion
Unsupervised Template-assisted Point Cloud Shape Correspondence Network
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 TANet, an unsupervised point cloud shape correspondence network that uses a learnable template bank (with a space aligner) and a template assistance module (template selector, correlation fusion, transitive consistency loss) to establish correspondences on TOSCA, SHREC'19, SMAL and SURREAL.
Problem
Unsupervised point cloud shape correspondence for non-rigid human and animal shapes with unconventional postures, where direct point-wise feature similarity is unreliable.
Approach
Learns a bank of explicitly structured point-cloud templates, adaptively selects a template per pair via geometric and semantic similarity, fuses template correlations into point features with attention, and enforces transitive consistency between template-mediated and direct similarity matrices.
Key results
- On TOSCA, TANet reaches 65.1% accuracy and 0.7 error, versus 52.3% and 1.2 for HSTR.
- On SHREC'19, TANet reaches 21.5% accuracy and 4.5 error, versus 19.3% and 4.9 for HSTR.
- On SMAL-to-TOSCA cross-dataset generalization, TANet reaches 37.1% accuracy, above SE-ORNet's 36.4%; on SURREAL-to-SHREC'19 it trails SE-ORNet (20.6% vs 21.5%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TANet instead of HSTR | point-cloud shape correspondence · TOSCA | correspondence accuracy (0.01): 52.3% -> 65.1% | Helped |
| use TANet instead of HSTR | point-cloud shape correspondence · SHREC’19 | correspondence accuracy (0.01): 19.3% -> 21.5% | Helped |
| add template generation and correlation fusion (ablation) | TANet · TOSCA | correspondence accuracy (0.01): 52.3% -> 54.6% | Helped |
| add template generation and transitive consistency loss (ablation) | TANet · TOSCA | correspondence accuracy (0.01): 52.3% -> 55.8% | Helped |
| add template construction loss (ablation) | [B] template generation and correlation fusion · TOSCA | correspondence accuracy (0.01): 54.6% -> 59.3% | Helped |
| add template construction loss (ablation) | [C] template generation and transitive consistency · TOSCA | correspondence accuracy (0.01): 55.8% -> 58.8% | Helped |
| add template selector (ablation) | [D] template construction loss and correlation fusion · TOSCA | correspondence accuracy (0.01): 59.3% -> 61.9% | Helped |
| add template selector (ablation) | [E] template construction loss and transitive consistency · TOSCA | correspondence accuracy (0.01): 58.8% -> 61.3% | Helped |
| add transitive consistency loss (ablation) | [F] template selector and correlation fusion · TOSCA | correspondence accuracy (0.01): 61.9% -> 65.1% | Helped |
| use 2 templates instead of 1 (ablation) | TANet · TOSCA | correspondence accuracy (0.01): 56.8% -> 61.4% | Helped |
| use 4 templates instead of 2 (ablation) | TANet · TOSCA | correspondence accuracy (0.01): 61.4% -> 65.1% | Helped |
| use 8 templates instead of 4 (ablation) | TANet · TOSCA | correspondence accuracy (0.01): 65.1% -> 66.2% | Helped |
Surprising findings
- Using template-assisted similarity at inference gives a small gain (66.3% vs 65.1% on TOSCA), but the authors use direct similarity at inference for efficiency.
Limitations
- Template-assisted similarity increases computational cost (FLOPs 4.29G vs 1.07G), so inference uses direct similarity instead.
- Explicit template structure remains sub-optimal without direct shape supervision, per the authors.
- Slightly lags SE-ORNet on SURREAL-to-SHREC'19 generalization, which is designed for symmetry issues.
Experimental setup
Datasets: TOSCA SHREC'19 SURREAL SMAL
Benchmarks: TOSCA SHREC'19 SMAL SURREAL SHREC'16 Owlii
Metrics: correspondence accuracy at 0.01 error tolerance average correspondence error
Baselines: SURFMNet CorrNet3D DPC SE-ORNet HSTR GeoFMNet Diff-FMaps 3D-CODED Elementary
Builds on: HSTR DPC SE-ORNet Chamfer distance Gumbel-softmax
Closest papers by meaning
- DiffCorr: Conditional Diffusion Model with Reliable Pseudo-Label Guidance for Unsupervised Point Cloud Shape Correspondence (AAAI 2025 · similarity 0.87)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.86)
- Adaptive Siamese Masked Autoencoder with Global Optimization for Unsupervised Point Cloud Shape Correspondence (AAAI 2025 · similarity 0.86)
- DV-Matcher: Deformation-based Non-rigid Point Cloud Matching Guided by Pre-trained Visual Features (CVPR 2025 · similarity 0.86)
- Orthogonal Dictionary Guided Shape Completion Network for Point Cloud (AAAI 2024 · similarity 0.86)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.86)
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.