Research map / Point clouds / Point cloud classification
Self-supervised point cloud pretraining
Pretrain point representations via cross-modal, contrastive and diffusion objectives.
73 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Point cloud classification, keyword cluster “language · pre · understanding”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 5 of 6 tracked benchmarks.
19 papers in 2023–24 and 42 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 8.2% to 16.2%.
Open questions the data points to
- ModelNet40 (accuracy): best went from 94.7 (up to 2024) to 96.2 in 2025 — GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · 0 citations)
- S3DIS Area 5 (mIoU): best went from 73.5 (up to 2024) to 74.3 in 2025 — How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need? (NeurIPS 2025 · 1 citation)
- ScanObjectNN OBJ-BG (accuracy): best went from 97.2 (up to 2024) to 98.97 in 2025 — GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · 0 citations)
- ModelNet40 (OA (%)): best went from 94.4 (up to 2024) to 96.89 in 2025 — PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning (AAAI 2025 · 26 citations)
- ScanObjectNN OBJ-ONLY (accuracy): best went from 91 (up to 2024) to 96.73 in 2025 — GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · 0 citations)
What looks settled
- ScanObjectNN PB-T50-RS (accuracy): best up to 2024 93.4, best in 2025–26 94.31 from 5 papers, less than 1% better.
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
- 2019 · PointDAN: A Multi-Scale 3D Domain Adaption Network for Point Cloud Representation (NeurIPS 2019 · 74 citations)
- 2020 · Self-Supervised Few-Shot Learning on Point Clouds (NeurIPS 2020 · 49 citations)
- 2021 · Self-Supervised Learning on 3D Point Clouds by Learning Discrete Generative Models (CVPR 2021 · 56 citations)
- 2022 · CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding (CVPR 2022 · 300 citations)
- 2023 · Self-Positioning Point-Based Transformer for Point Cloud Understanding (CVPR 2023 · 123 citations)
- 2024 · PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · 81 citations)
- 2025 · Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · 82 citations)
- 2026 · RadarLLM: Empowering Large Language Models to Understand Human Motion from Millimeter-wave Point Cloud Sequence (AAAI 2026 · 4 citations)
Earliest papers here
- PointDAN: A Multi-Scale 3D Domain Adaption Network for Point Cloud Representation (NeurIPS 2019 · 74 citations)
- Self-Supervised Deep Learning on Point Clouds by Reconstructing Space (NeurIPS 2019 · 33 citations)
- Self-Supervised Few-Shot Learning on Point Clouds (NeurIPS 2020 · 49 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 PB-T50-RS · accuracy (higher is better)
8 papers, 2022–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 89.3 | P2P: Tuning Pre-trained Image Models for Point Cloud Analysis with Point-to-Pixel Prompting (NeurIPS 2022 · 36 citations) |
| 2023 | 93.4 | PointGPT: Auto-regressively Generative Pre-training from Point Clouds (NeurIPS 2023 · 33 citations) |
| 2025 | 94.31 | GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · 0 citations) |
ModelNet40 · accuracy (higher is better)
7 papers, 2019–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 92.4 | Self-Supervised Deep Learning on Point Clouds by Reconstructing Space (NeurIPS 2019 · 33 citations) |
| 2022 | 94 | P2P: Tuning Pre-trained Image Models for Point Cloud Analysis with Point-to-Pixel Prompting (NeurIPS 2022 · 36 citations) |
| 2024 | 94.7 | PointPatchMix: Point Cloud Mixing with Patch Scoring (AAAI 2024 · 10 citations) |
| 2025 | 96.2 | GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · 0 citations) |
S3DIS Area 5 · mIoU (higher is better)
7 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 71.3 | A Unified Query-Based Paradigm for Point Cloud Understanding (CVPR 2022 · 52 citations) |
| 2023 | 73.5 | ConDaFormer: Disassembled Transformer with Local Structure Enhancement for 3D Point Cloud Understanding (NeurIPS 2023 · 9 citations) |
| 2025 | 74.3 | How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need? (NeurIPS 2025 · 1 citation) |
ScanObjectNN OBJ-BG · accuracy (higher is better)
7 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 97.2 | PointGPT: Auto-regressively Generative Pre-training from Point Clouds (NeurIPS 2023 · 33 citations) |
| 2025 | 98.97 | GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · 0 citations) |
ModelNet40 · OA (%) (higher is better)
6 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 94.4 | Let Images Give You More: Point Cloud Cross-Modal Training for Shape Analysis (NeurIPS 2022 · 2 citations) |
| 2025 | 96.89 | PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning (AAAI 2025 · 26 citations) |
ScanObjectNN OBJ-ONLY · accuracy (higher is better)
6 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 91 | PointPatchMix: Point Cloud Mixing with Patch Scoring (AAAI 2024 · 10 citations) |
| 2025 | 96.73 | GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · 0 citations) |
What has been tried
54 papers here report 646 measured changes (a component added, removed or swapped, compared with a baseline): helped 76%, no clear effect 4%, hurt 19%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use multiple serialization strategies instead of a single strategy | ScanObjectNN · OA | 0.42 OA | Helped | Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · 82 citations) |
| Use Hilbert and Trans-Hilbert scanning instead of random serialization | ScanObjectNN OBJ-BG · overall accuracy | +1.20% | Helped | PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · 81 citations) |
| add frozen pre-trained T5 instead of the Text2Pos LSTM | KITTI360Pose test · Submap retrieval recall, k=1 | 0.26 -> 0.28 | Helped | Text2Loc: 3D Point Cloud Localization from Natural Language (CVPR 2024 · 46 citations) |
| Replace the point condition network with pointwise concatenation | S3DIS Area 5 · mIoU (%) | 70.02 -> 69.43 | Hurt | Point Cloud Pre-training with Diffusion Models (CVPR 2024 · 41 citations) |
| Add TFTS | ScanObjectNN PB T50 RS · overall accuracy | 75.99 -> 83.17 | Helped | Dynamic Adapter Meets Prompt Tuning: Parameter-Efficient Transfer Learning for Point Cloud Analysis (CVPR 2024 · 35 citations) |
| add contrastive pre-training | SJTU-PCQA · SROCC | 0.806 -> 0.897 | Helped | Contrastive Pre-Training with Multi-View Fusion for No-Reference Point Cloud Quality Assessment (CVPR 2024 · 27 citations) |
| Add BQE | ScanObjectNN PB-T50-RS · OA | 89.6 -> 90.4 | Helped | PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning (AAAI 2025 · 26 citations) |
| remove the hierarchical decoder during self-supervised learning | ScanNet semantic segmentation · mIoU | 20.7% to 60.4% | Helped | Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · 19 citations) |
| combine color, depth, and text modalities rather than use color and depth only | LS-PCQA · PLCC | 0.688 -> 0.755 | Helped | CLIP-PCQA: Exploring Subjective-Aligned Vision-Language Modeling for Point Cloud Quality Assessment (AAAI 2025 · 11 citations) |
| Replace block-level mixing with patch-level mixing | ModelNet40 · accuracy | 91.7 -> 92.4 | Helped | PointPatchMix: Point Cloud Mixing with Patch Scoring (AAAI 2024 · 10 citations) |
| Remove all Mamba modules | ScanNet200 Val · mIoU | 36.3 -> 26.1 | Hurt | Pamba: Enhancing Global Interaction in Point Clouds via State Space Model (AAAI 2025 · 9 citations) |
| add DAF-Conv to EdgeConv + MHA | ScanObjectNN PB_T50_RS · OA | 91.3 -> 96.7 | Helped | Point Clouds Meets Physics: Dynamic Acoustic Field Fitting Network for Point Cloud Understanding (CVPR 2025 · 8 citations) |
What reviewers will expect
Datasets and baselines named in the 27 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 · 10 of 27 papers
- ScanObjectNN · 10 of 27 papers
- ShapeNetPart · 9 of 27 papers
- S3DIS · 8 of 27 papers
- ShapeNet · 8 of 27 papers
- ScanNet · 6 of 27 papers
- ScanNet200 · 3 of 27 papers
- nuScenes · 3 of 27 papers
Baselines
- Point-MAE · 8 of 27 papers
- PointMamba · 5 of 27 papers
- Point-BERT · 5 of 27 papers
- ReCon · 5 of 27 papers
- PointNeXt · 4 of 27 papers
- IDPT · 4 of 27 papers
- DAPT · 4 of 27 papers
- PTv3 · 3 of 27 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 61 papers; from OpenAlex and the papers' affiliations).
- Shanghai Jiao Tong University · 6 papers
- Nanyang Technological University · 6 papers
- Zhejiang University · 4 papers
- Peking University · 3 papers
- University of Science and Technology of China · 3 papers
- Tencent · 3 papers
- Tsinghua University · 3 papers
- Shanghai AI Laboratory · 3 papers
- National University of Singapore · 2 papers
- Zhejiang Lab · 2 papers
Use it
Download BibTeX of 33 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)
- Masked point autoencoders (22 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).