Research map / Point clouds / Point cloud classification
Masked point autoencoders
Pretrain point clouds by reconstructing masked patches, often multi-scale or rotation-invariant.
22 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Point cloud classification, keyword cluster “language · pre · understanding”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
10 papers in 2023–24 and 8 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 4.3% to 3.1%.
No longer reported by papers here: ScanObjectNN OBJ-BG (accuracy, last in 2024); ScanObjectNN OBJ-ONLY (accuracy, last in 2024); ScanObjectNN PB-T50-RS (accuracy, last in 2024).
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2018 · Representer Point Selection for Explaining Deep Neural Networks (NeurIPS 2018 · 70 citations)
- 2021 · TearingNet: Point Cloud Autoencoder To Learn Topology-Friendly Representations (CVPR 2021 · 36 citations)
- 2022 · Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling (CVPR 2022 · 780 citations)
- 2023 · Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · 128 citations)
- 2024 · Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · 34 citations)
- 2025 · Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · 12 citations)
- 2026 · Point-SRA: Self-Representation Alignment for 3D Representation Learning (AAAI 2026 · 2 citations)
Earliest papers here
- Representer Point Selection for Explaining Deep Neural Networks (NeurIPS 2018 · 70 citations)
- TearingNet: Point Cloud Autoencoder To Learn Topology-Friendly Representations (CVPR 2021 · 36 citations)
- Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling (CVPR 2022 · 780 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
ScanObjectNN OBJ-BG · accuracy (higher is better)
8 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 91.22 | Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training (NeurIPS 2022 · 117 citations) |
| 2023 | 94.15 | Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · 128 citations) |
| 2024 | 95.52 | PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · 7 citations) |
ScanObjectNN OBJ-ONLY · accuracy (higher is better)
8 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 88.81 | Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training (NeurIPS 2022 · 117 citations) |
| 2023 | 91.57 | Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · 128 citations) |
| 2024 | 94.32 | PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · 7 citations) |
ScanObjectNN PB-T50-RS · accuracy (higher is better)
8 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 86.43 | Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training (NeurIPS 2022 · 117 citations) |
| 2023 | 90.11 | Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · 128 citations) |
| 2024 | 90.35 | PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · 7 citations) |
ModelNet40 10-way 10-shot · accuracy (higher is better)
6 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 91 | Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling (CVPR 2022 · 780 citations) |
| 2023 | 92.6 | Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · 128 citations) |
| 2024 | 94 | Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · 34 citations) |
ModelNet40 10-way 20-shot · accuracy (higher is better)
6 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 92.7 | Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling (CVPR 2022 · 780 citations) |
| 2023 | 95.5 | Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · 128 citations) |
| 2024 | 95.9 | PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · 7 citations) |
ModelNet40 5-way 10-shot · accuracy (higher is better)
6 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 94.6 | Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling (CVPR 2022 · 780 citations) |
| 2023 | 97 | Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · 128 citations) |
| 2024 | 97.4 | PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · 7 citations) |
What has been tried
18 papers here report 208 measured changes (a component added, removed or swapped, compared with a baseline): helped 75%, no clear effect 3%, hurt 19%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add hybrid global and local masking to Point-MAE | ScanObjectNN OBJ-BG · accuracy | 90.02 -> 89.67 | Hurt | Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · 34 citations) |
| Use SAST instead of HLT for segmentation | ShapeNetPart, training from pretrained · mIoU (Inst.) | 86.1 -> 85.7 | Hurt | Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · 12 citations) |
| add autoregressive mask generation to autoencoding | ModelNet40 · accuracy | 92.9 -> 93.8 | Helped | General Point Model Pretraining with Autoencoding and Autoregressive (CVPR 2024 · 12 citations) |
| Add the dual-branch self-supervised learning strategy to RI-Transformer | ScanObjectNN OBJ_BG z/SO3 · accuracy | 88.3 -> 91.6 | Helped | RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning (AAAI 2025 · 9 citations) |
| add center prediction with PCM to reconstruction pre-training | ScanObjectNN OBJ-BG · accuracy | 92.94 -> 94.32 | Helped | PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · 7 citations) |
| Use normals and surface variation together instead of either feature alone | ScanObjectNN · accuracy | 86.5, 87.0, 87.7 | Helped | 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · 5 citations) |
| add local aggregation to an encoder with a simple two-layer MLP | ScanObjectNN · classification performance | Helped | LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling (NeurIPS 2024 · 2 citations) | |
| Combine the global CLS and masked point modeling objectives instead of training either objective alone | ScanObjectNN PB_T50_RS, full fine-tune · accuracy | 88.72 -> 90.53; 88.58 -> 90.53 | Helped | Asymmetric Dual Self-Distillation for 3D Self-Supervised Representation Learning (NeurIPS 2025 · 0 citations) |
| Use single-task rather than multi-task fine-tuning | GEOM-DRUGS conformer generation · Recall COV Avg | 50.1 -> 57.7 | Helped | nach0-pc: Multi-task Language Model with Molecular Point Cloud Encoder (AAAI 2025) |
What reviewers will expect
Datasets and baselines named in the 6 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ModelNet40 · 4 of 6 papers
- ShapeNet · 3 of 6 papers
- ScanObjectNN · 3 of 6 papers
- ShapeNetPart · 3 of 6 papers
- S3DIS · 2 of 6 papers
Baselines
- Point-BERT · 3 of 6 papers
- Point-MAE · 3 of 6 papers
- Point-M2AE · 3 of 6 papers
- ACT · 2 of 6 papers
- PointGPT-S · 2 of 6 papers
- PointNet++ · 2 of 6 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 18 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 4 papers
- Shanghai Jiao Tong University · 3 papers
- Shenzhen University · 2 papers
- Harbin Institute of Technology · 2 papers
- Peking University · 2 papers
- Sun Yat-sen University · 2 papers
Use it
Download BibTeX of 22 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Point-language and open-vocabulary 3D (107 papers)
- Self-supervised point cloud pretraining (73 papers)
Elsewhere in the topic
- Local feature aggregation networks (158 papers)
- Scene flow and dynamic point clouds (89 papers)
- Equivariant point cloud networks (84 papers)
- LiDAR point cloud compression and completion (60 papers)
- Point cloud detection and tracking (56 papers)
- Efficient point sampling architectures (45 papers)
- Adversarial point cloud attacks (37 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).