Research map / Point clouds / Point cloud classification
DV-Matcher: Deformation-based Non-rigid Point Cloud Matching Guided by Pre-trained Visual Features
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 DV-Matcher, a label-free point cloud non-rigid matching framework that injects DINOv2 (FeatUp-upsampled) multi-view visual features into point features and trains with a single-feedforward deformation-based proxy loss plus ARAP, smoothness and geometric similarity losses.
Problem
Dense correspondence estimation between non-rigidly deformable point clouds without correspondence labels, meshes, or slow iterative optimization, robust to pose/style change, partiality, density and noise.
Approach
Aggregate pixel-level pre-trained 2D vision features (DINOv2 with FeatUp) from multi-view projections back to 3D as visual encoding, and train a feedforward network to predict deformation graph transformations so that correspondence-induced deformation aligns the shapes, replacing reconstruction-based proxy losses that cause mode collapse.
Key results
- On SCAPE_r/SHREC19_r cross-dataset setting, DV-Matcher achieves 7.2 mean geodesic error versus 23.6 for the best pure point-cloud prior method (SE-ORNet).
- On SHREC'16 Cuts and Holes partial matching, error 16.9 and 13.0 versus 32.9 and 27.6 for the best pure point-cloud baselines (DPC and SE-ORNet).
- Average matching time is 3.22 s per shape on SCAPE_r, slower than DPC (1.41 s) and SE-ORNet (0.85 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add visual encoding from pre-trained vision features (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 15.1 -> 6.2 | Helped |
| add deformation-based loss (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 11.2 -> 6.2 | Helped |
| add geometric similarity loss (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 12.8 -> 6.2 | Helped |
| add smoothness loss (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 6.2 -> 6.9 | Hurt |
| add LG-Net (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 17.3 -> 6.2 | Helped |
| add FeatUp (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 7.8 -> 6.2 | Helped |
| add local attention network (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 8.0 -> 6.2 | Helped |
| add global attention network (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 8.1 -> 6.2 | Helped |
| add positional encoding (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 7.1 -> 6.2 | Helped |
| add fusion module (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 7.9 -> 6.2 | Helped |
| add reconstruction loss (ablation) | DV-Matcher · SCAPE_r | mean geodesic error (×100): 6.2 -> 7.6 | Hurt |
| train DV-Matcher rather than the strongest listed baseline | non-rigid point cloud matching · SCAPE_r/SHREC’19_r (train SCAPE_r) | mean geodesic error (×100): 23.6 -> 7.2 | Helped |
Surprising findings
- Adding a reconstruction loss causes shape collapse, whereas the deformation-based loss avoids it.
- Reconstruction-based baselines DPC and SE-ORNet degrade by more than 8% when tested on denser point clouds, whereas DV-Matcher degrades by only 0.46%.
Limitations
- Inference takes about 3.22 s per shape, far from real-time.
- Does not guarantee robustness to arbitrary orientation of inputs.
- Difficult to train directly on raw scans without overlap region prediction.
Experimental setup
Datasets: SCAPE_r FAUST_r SHREC'19_r DT4D-H SHREC'07-H SHREC'16 SURREAL TOSCA SMAL GarmCap Spleen Pancreas
Benchmarks: SCAPE_r FAUST_r SHREC'19_r DT4D-H SHREC'07-H SCAPE-PV FAUST-PV SHREC'16 Cuts/Holes SHREC'19 partial TOSCA Spleen Pancreas
Metrics: mean geodesic error (x100) chamfer distance earth mover's distance
Baselines: DPC SE-ORNet SSMSM NIE DiffFMaps SyNoRiM CorrNet3D TransMatch DPFM PointSetReg Point2SSM HCLV2S
Builds on: I2P-MAE DINOv2 FeatUp DPC NIE Point Transformer (PCT)
Closest papers by meaning
- Lepard: Learning Partial Point Cloud Matching in Rigid and Deformable Scenes (CVPR 2022 · similarity 0.89)
- APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction (NeurIPS 2025 · similarity 0.88)
- Dataset Distillation of 3D Point Clouds via Distribution Matching (NeurIPS 2025 · similarity 0.87)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.87)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · 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.