Research map / Point clouds / Shape and surface completion
DiffCorr: Conditional Diffusion Model with Reliable Pseudo-Label Guidance for Unsupervised Point Cloud Shape Correspondence
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 DiffCorr, a conditional DDIM-style diffusion model with a transformer backbone and a reliable pseudo-label generator (optimal transport plus cycle-consistency) for unsupervised point cloud shape correspondence on human and animal datasets.
Problem
Unsupervised dense point cloud correspondence between non-rigid human and animal shapes fails under large-scale motion displacements when using one-step similarity-based matching, and no ground-truth labels guide refinement.
Approach
Replace one-step matching with coarse-to-fine multi-step denoising of flow residuals, conditioned on an initial flow from embedding similarity and a DGCNN-encoded local cost, with pseudo-flow supervision from Sinkhorn optimal transport filtered by cycle-consistent matching and variance-weighted denoising loss.
Key results
- On TOSCA/TOSCA, DiffCorr reaches 66.71% accuracy and 0.65 error versus 65.1% and 0.7 for the best prior method TANet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiffCorr instead of SE-ORNet | point cloud shape correspondence · SURREAL/SHREC’19 | accuracy: 21.5% -> 22.53±0.27% | Helped |
| use DiffCorr instead of SE-ORNet | point cloud shape correspondence · SURREAL/SHREC’19 | err: 4.6 -> 4.31±0.08 | Helped |
| use DiffCorr instead of TANet | point cloud shape correspondence · SHREC’19/SHREC’19 | accuracy: 21.5% -> 22.16±0.35% | Helped |
| use DiffCorr instead of TANet | point cloud shape correspondence · SHREC’19/SHREC’19 | err: 4.5 -> 4.36±0.07 | Helped |
| use DiffCorr instead of TANet | point cloud shape correspondence · SMAL/TOSCA | accuracy: 37.1% -> 41.64 ± 0.56% | Helped |
| use DiffCorr instead of TANet | point cloud shape correspondence · SMAL/TOSCA | err: 3.7 -> 2.31± 0.12 | Helped |
| use DiffCorr instead of TANet | point cloud shape correspondence · TOSCA/TOSCA | accuracy: 65.1% -> 66.71±1.79% | Helped |
| use DiffCorr instead of TANet | point cloud shape correspondence · TOSCA/TOSCA | err: 0.7 -> 0.65 ± 0.03 | Helped |
| add T-CDM, local cost, initial flow, optimal transport, and cross-consistent matching (ablation) | DiffCorr · TOSCA/TOSCA | accuracy: 54.58% -> 66.71% | Helped |
| add local cost, initial flow, and cross-consistent matching (ablation) | DiffCorr with T-CDM and optimal transport · TOSCA/TOSCA | accuracy: 60.82% -> 66.71% | Helped |
| add local cost, initial flow, and optimal transport (ablation) | DiffCorr with T-CDM and cross-consistent matching · TOSCA/TOSCA | accuracy: 60.15% -> 66.71% | Helped |
| add local cost and initial flow (ablation) | DiffCorr with T-CDM, optimal transport, and cross-consistent matching · TOSCA/TOSCA | accuracy: 62.79% -> 66.71% | Helped |
Surprising findings
- Hard optimal-transport labels (epsilon=0) gave suboptimal or even degraded performance, so Sinkhorn with entropic regularization was used instead.
Limitations
- Slow inference and high randomness inherent to diffusion models
- Background elements in real scans increase failure cases
Experimental setup
Datasets: SURREAL SHREC'19 SMAL TOSCA
Benchmarks: SHREC'19 TOSCA SURREAL/SHREC'19 SMAL/TOSCA
Metrics: correspondence accuracy at 0.01 tolerance average correspondence error
Baselines: Diff-FMaps 3D-CODED Elementary CorrNet3D DPC HSTR SE-ORNet TANet
Builds on: DDIM DPC HSTR SE-ORNet adaLN-Zero
Closest papers by meaning
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.90)
- Where Precision Meets Efficiency: Transformation Diffusion Model for Point Cloud Registration (AAAI 2025 · similarity 0.88)
- Diffusion 3D Features (Diff3F): Decorating Untextured Shapes with Distilled Semantic Features (CVPR 2024 · similarity 0.87)
- Unsupervised Template-assisted Point Cloud Shape Correspondence Network (CVPR 2024 · similarity 0.87)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.87)
- Adaptive Siamese Masked Autoencoder with Global Optimization for Unsupervised Point Cloud Shape Correspondence (AAAI 2025 · similarity 0.87)
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.