Research map / Point clouds / Shape and surface completion
Adaptive Siamese Masked Autoencoder with Global Optimization 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 AMIGO, an unsupervised point cloud shape correspondence method combining a downsampled siamese masked autoencoder with adversarial adaptive masking and optimal-transport pseudo-label global optimization used only in training.
Problem
Unsupervised point cloud shape correspondence is slow because existing methods encode all dense points, and independent per-point matching yields multiple-to-one erroneous correspondences.
Approach
Downsample the point clouds 8x and train a siamese masked autoencoder where masked source regions are reconstructed using target point features, with an adversarially learned mask generator; optimal transport (Sinkhorn) produces global pseudo-labels for a cross-entropy loss during training only, adding no inference cost.
Key results
- On SHREC'19 intra-dataset, AMIGO reaches 22.7% accuracy and 4.3 error, versus 21.5% and 4.5 for TANet.
- On TOSCA intra-dataset, AMIGO reaches 65.9% accuracy and 0.5 error, versus 65.1% and 0.7 for TANet.
- Ablation on SHREC'19 shows global optimization adds about 1.1% accuracy on top of ASMAE and AAMG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add downsampling (ablation) | AMIGO · SHREC’19 | accuracy: 19.3% -> 16.5% | Hurt |
| add ASMAE to the downsampled model (ablation) | AMIGO · SHREC’19 | accuracy: 16.5% -> 20.9% | Helped |
| add AAMG to DS + ASMAE (ablation) | AMIGO · SHREC’19 | accuracy: 20.9% -> 21.6% | Helped |
| add GO to DS + ASMAE + AAMG (ablation) | AMIGO · SHREC’19 | accuracy: 21.6% -> 22.7% | Helped |
| add GO to DS + ASMAE without AAMG (ablation) | AMIGO · SHREC’19 | accuracy: 20.9% -> 21.9% | Helped |
| use AAMG instead of random masks at a 20% ratio (ablation) | ASMAE · SHREC’19 | acc(0.01): 19.29% -> 19.85% | Helped |
| use AAMG instead of random masks at a 40% ratio (ablation) | ASMAE · SHREC’19 | acc(0.01): 21.86% -> 22.72% | Helped |
| use AAMG instead of random masks at a 60% ratio (ablation) | ASMAE · SHREC’19 | acc(0.01): 21.31% -> 22.16% | Helped |
| use AAMG instead of random masks at an 80% ratio (ablation) | ASMAE · SHREC’19 | acc(0.01): 20.67% -> 21.52% | Helped |
| vary AAMG mask ratio from 40% to 20% (ablation) | ASMAE · SHREC’19 | acc(0.01): 22.72% -> 19.85% | Hurt |
| vary AAMG mask ratio from 40% to 60% (ablation) | ASMAE · SHREC’19 | acc(0.01): 22.72% -> 22.16% | Hurt |
| vary AAMG mask ratio from 40% to 80% (ablation) | ASMAE · SHREC’19 | acc(0.01): 22.72% -> 21.52% | Hurt |
Surprising findings
- Naive downsampling without the masked autoencoder reduces accuracy from 19.3% to 16.5% on SHREC'19.
Limitations
- Gains over the prior state of the art are small on SHREC'19 and SURREAL (around 1-2 accuracy points).
- Inference speed improvement is claimed but not quantified in the main text.
Experimental setup
Datasets: SURREAL SHREC'19 SMAL TOSCA
Benchmarks: SHREC'19 SURREAL TOSCA SMAL
Metrics: correspondence accuracy at 0.01 tolerance average correspondence error
Baselines: SURFMNet GeoFMNet Diff-FMaps 3D-CODED Elementary CorrNet3D DPC SE-ORNet HSTR TANet
Builds on: Point-MAE HSTR DPC Sinkhorn optimal transport Gumbel-softmax
Closest papers by meaning
- RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning (AAAI 2025 · similarity 0.89)
- 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · similarity 0.88)
- SIEVE: Overcoming Topological Obstruction in Equivariant Self-Supervised Learning (NeurIPS 2026 · similarity 0.87)
- DiffCorr: Conditional Diffusion Model with Reliable Pseudo-Label Guidance for Unsupervised Point Cloud Shape Correspondence (AAAI 2025 · similarity 0.87)
- Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · similarity 0.87)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (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.