Research map / Point clouds / Point cloud classification
RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning
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 RI-MAE, a rotation-invariant masked point modeling framework with a PCA-based RI-Transformer backbone and a dual-branch student-teacher design that reconstructs masked patch latents rather than raw coordinates.
Problem
Masked point modeling methods for point clouds are sensitive to rotations, causing sharp performance drops on arbitrarily rotated inputs; reconstructing raw coordinates also gives inconsistent supervision under rotation.
Approach
Decompose each PCA-aligned point patch into rotation-invariant canonical content and orientation; use rotation-invariant relative orientation (R_j R_i^T) as attention bias and rotation-invariant position embedding (c_i R_i^T); and perform masked-patch prediction in a latent space via an EMA teacher-student dual-branch with an RI-Transformer predictor.
Key results
- RI-MAE reaches 91.6% on ScanObjectNN OBJ_BG in z/z, SO3/SO3 and z/SO3 settings, versus 86.6% for the best prior rotation-invariant method reported.
- On ShapeNetPart z/SO3, RI-MAE achieves 84.3 mIoU_I and 82.1 mIoU_C, while Point-MAE-style MPM baselines reach around 31-36 mIoU_I.
- Dual-branch design improves ScanObjectNN OBJ_BG z/SO3 accuracy from 86.7% (AE variant) to 91.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the dual-branch self-supervised learning strategy to RI-Transformer (ablation) | RI-Transformer · ScanObjectNN OBJ_BG z/SO3 | accuracy: 88.3 -> 91.6 | Helped |
| Add RI-OE while retaining the dual-branch architecture and omitting RI-PE (ablation) | RI-MAE · ScanObjectNN OBJ_BG z/SO3 | accuracy: 82.6 -> 88.6 | Helped |
| Add RI-PE while retaining the dual-branch architecture and omitting RI-OE (ablation) | RI-MAE · ScanObjectNN OBJ_BG z/SO3 | accuracy: 82.6 -> 88.5 | Helped |
| Add RI-PE to the dual-branch method with RI-OE (ablation) | RI-MAE · ScanObjectNN OBJ_BG z/SO3 | accuracy: 88.6 -> 91.6 | Helped |
| Add RI-OE to the dual-branch method with RI-PE (ablation) | RI-MAE · ScanObjectNN OBJ_BG z/SO3 | accuracy: 88.5 -> 91.6 | Helped |
| Replace the autoencoder with the dual-branch architecture (ablation) | RI-MAE · ScanObjectNN OBJ_BG z/SO3 | accuracy: 86.7 -> 91.6 | Helped |
| Use RI-MAE rather than RI-Transformer pretraining | RI-Transformer · ModelNet40 10-way 10-shot z/SO3 | accuracy: 78.9 ± 6.4 -> 90.2 ± 5.5 | Helped |
| Use RI-MAE rather than RI-Transformer pretraining | RI-Transformer · ModelNet40 10-way 20-shot z/SO3 | accuracy: 86.8 ± 5.2 -> 93.7 ± 3.5 | Helped |
| Use RI-MAE rather than RI-Transformer | RI-Transformer · ShapeNetPart z/SO3 | mIoU_I: 84.0 -> 84.3 | Helped |
| Use RI-MAE rather than RI-Transformer | RI-Transformer · ShapeNetPart z/SO3 | mIoU_C: 81.6 -> 82.1 | Helped |
| Use RI-MAE rather than RI-Transformer | RI-Transformer · S3DIS Area 5 z/SO3 | mIoU: 59.9 -> 60.3 | Helped |
Surprising findings
- Standard MPM baselines drop from roughly 80-90% to about 20-25% on ScanObjectNN under z/SO3 testing, showing extreme rotation sensitivity.
- Under z/z on ModelNet40 few-shot, the RI-Transformer baseline and RI-MAE are below Point-MAE and ACT, indicating that the rotation-invariant design trades off some in-distribution accuracy.
Limitations
- Relies on PCA for canonical pose, which can be ambiguous and sign-unstable (not discussed in the text)
- Rotation-invariance is evaluated mainly under azimuthal and SO(3) rotation scenarios on a few benchmarks; no robustness analysis against noise or occlusion is reported
Experimental setup
Datasets: ShapeNet ScanObjectNN ModelNet40 ShapeNetPart S3DIS
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart S3DIS
Metrics: accuracy mIoU few-shot accuracy
Baselines: Point-BERT Point-MAE Point-M2AE ACT PointGPT-S Transformer RaRI-Conv CRIN LocoTrans CrossPoint Yu et al.
Builds on: Point-MAE MAE BERT ClusterNet RI-Conv++
Closest papers by meaning
- 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · similarity 0.90)
- PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · similarity 0.89)
- Adaptive Siamese Masked Autoencoder with Global Optimization for Unsupervised Point Cloud Shape Correspondence (AAAI 2025 · similarity 0.89)
- SIEVE: Overcoming Topological Obstruction in Equivariant Self-Supervised Learning (NeurIPS 2026 · similarity 0.88)
- Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · similarity 0.88)
- LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling (NeurIPS 2024 · 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.