Research map / Point clouds / Point cloud classification
Mitigating Positional Leakage in 3D Masked Autoencoders for Robust Representation Learning
Xu Yan, Huiqun Wang, Chen Wang, Lei Ren, Di Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add recalibrated positional embedding (RPE) (ablation) | Point-MAE baseline · ScanObjectNN PB-T50 | accuracy (%): 88.58 -> 89.41 | Improved |
| add gated positional interface (GPI) (ablation) | Point-MAE baseline · ScanObjectNN PB-T50 | accuracy (%): 88.58 -> 89.31 | Improved |
| combine RPE and GPI (ablation) | MPL-MAE · ScanObjectNN PB-T50 | accuracy (%): 88.58 -> 90.35 | Improved |
| add topology preservation loss L_topo (ablation) | MPL-MAE · ScanObjectNN PB-T50 | accuracy (%): 89.31 -> 90.35 | Improved |
| use cosine instead of Euclidean distance in RPE (ablation) | MPL-MAE · ScanObjectNN PB-T50 | accuracy (%): 89.27 -> 90.35 | Improved |
| use Gumbel-Softmax instead of Sigmoid gating in GPI (ablation) | MPL-MAE · ScanObjectNN PB-T50 | accuracy (%): 89.03 -> 90.35 | Improved |
| apply MPL-MAE over Point-MAE with pre-training augmentation removed (ablation) | MPL-MAE · ModelNet40 linear probe | accuracy (%): 89.5 -> 92.2 (w/o Aug); 90.6 -> 92.9 (w/ Aug) | Improved |
| replace Point-MAE with MPL-MAE under Gaussian noise | MPL-MAE · ModelNet40 noise sigma 3.0 | accuracy (%): 75.6 -> 80.8 | Improved |
| replace PCP-MAE with MPL-MAE in full fine-tuning | MPL-MAE · ModelNet40 w/ vote | accuracy (%): 94.2 -> 94.4 (Point-FEMAE 94.5 higher) | Mixed |
| replace PCP-MAE with MPL-MAE for segmentation | MPL-MAE · S3DIS | mIoU: 61.3 -> 61.1 (mAcc 71.0 -> 71.7) | Mixed |
| use MPL-MAE pretrained encoder for registration | MPL-MAE · ModelNet40 registration | MSE(R): 0.583 -> 0.335 (vs Point-MAE) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PCP-MAE — PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024) · better on 7 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · similarity 0.90)
- Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · similarity 0.89)
- RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning (AAAI 2025 · similarity 0.88)
- PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · similarity 0.87)
- SIEVE: Overcoming Topological Obstruction in Equivariant Self-Supervised Learning (NeurIPS 2026 · similarity 0.87)
- Multi-Scale Neighborhood Occupancy Masked Autoencoder for Self-Supervised Learning in LiDAR Point Clouds (CVPR 2025 · similarity 0.85)
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.
