Research map / Point clouds / Point cloud classification
Local feature aggregation networks
Design point-set convolutions and feature extractors for point cloud recognition.
158 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Point cloud classification, keyword cluster “adversarial · convolution · invariant”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 6 tracked benchmarks.
39 papers in 2023–24 and 34 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 16.8% to 13.1%.
Open questions the data points to
- S3DIS Area 5 (mIoU): best went from 74.3 (up to 2024) to 75.7 in 2025 — DeepLA-Net: Very Deep Local Aggregation Networks for Point Cloud Analysis (CVPR 2025 · 5 citations)
- Papers report Baseline on S3DIS Area 5 (mIoU) anywhere from 55.88 to 70.9: comparisons against it depend on whose number you take.
- Papers report PointNet++ on S3DIS (mIoU (%)) anywhere from 47.6 to 56.36: comparisons against it depend on whose number you take.
No longer reported by papers here: ModelNet40 (accuracy, last in 2023); ShapeNetPart (instance mIoU, last in 2024); ModelNet40 z/z (accuracy, last in 2024); ModelNet40 z/SO(3) (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
- 2016 · Contour Detection in Unstructured 3D Point Clouds (CVPR 2016 · 226 citations)
- 2017 · PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation (CVPR 2017 · 9,885 citations)
- 2018 · SO-Net: Self-Organizing Network for Point Cloud Analysis (CVPR 2018 · 986 citations)
- 2019 · PointConv: Deep Convolutional Networks on 3D Point Clouds (CVPR 2019 · 1,944 citations)
- 2020 · PointASNL: Robust Point Clouds Processing Using Nonlocal Neural Networks With Adaptive Sampling (CVPR 2020 · 677 citations)
- 2021 · PAConv: Position Adaptive Convolution With Dynamic Kernel Assembling on Point Clouds (CVPR 2021 · 517 citations)
- 2022 · Stratified Transformer for 3D Point Cloud Segmentation (CVPR 2022 · 557 citations)
- 2023 · Starting From Non-Parametric Networks for 3D Point Cloud Analysis (CVPR 2023 · 121 citations)
- 2024 · Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · 600 citations)
- 2025 · LidarGait++: Learning Local Features and Size Awareness from LiDAR Point Clouds for 3D Gait Recognition (CVPR 2025 · 9 citations)
- 2026 · Hierarchical Direction Perception via Atomic Dot-Product Operators for Rotation-Invariant Point Clouds Learning (AAAI 2026 · 3 citations)
Earliest papers here
- Contour Detection in Unstructured 3D Point Clouds (CVPR 2016 · 226 citations)
- Learning to Assign Orientations to Feature Points (CVPR 2016 · 121 citations)
- Accelerated Generative Models for 3D Point Cloud Data (CVPR 2016 · 70 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.
S3DIS Area 5 · mIoU (higher is better)
18 papers, 2019–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 60.28 | PointWeb: Enhancing Local Neighborhood Features for Point Cloud Processing (CVPR 2019 · 886 citations) |
| 2020 | 63.6 | SegGCN: Efficient 3D Point Cloud Segmentation With Fuzzy Spherical Kernel (CVPR 2020 · 120 citations) |
| 2021 | 66.8 | ASSANet: An Anisotropic Separable Set Abstraction for Efficient Point Cloud Representation Learning (NeurIPS 2021 · 39 citations) |
| 2022 | 72 | Stratified Transformer for 3D Point Cloud Segmentation (CVPR 2022 · 557 citations) |
| 2023 | 73.3 | Improving Graph Representation for Point Cloud Segmentation via Attentive Filtering (CVPR 2023 · 76 citations) |
| 2024 | 74.3 | X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · 19 citations) |
| 2025 | 75.7 | DeepLA-Net: Very Deep Local Aggregation Networks for Point Cloud Analysis (CVPR 2025 · 5 citations) |
ModelNet40 · accuracy (higher is better)
14 papers, 2018–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 91 | Mining Point Cloud Local Structures by Kernel Correlation and Graph Pooling (CVPR 2018 · 578 citations) |
| 2019 | 93.6 | Relation-Shape Convolutional Neural Network for Point Cloud Analysis (CVPR 2019 · 1,003 citations) |
| 2021 | 93.9 | PAConv: Position Adaptive Convolution With Dynamic Kernel Assembling on Point Clouds (CVPR 2021 · 517 citations) |
ModelNet40 · OA (higher is better)
9 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 92.6 | A-CNN: Annularly Convolutional Neural Networks on Point Clouds (CVPR 2019 · 292 citations) |
| 2022 | 94.5 | Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework (ICLR 2022 · 309 citations) |
| 2023 | 94.7 | GAM: Gradient Attention Module of Optimization for Point Clouds Analysis (AAAI 2023 · 21 citations) |
| 2025 | 94.8 | Dual Manifold Regularization Steered Robust Representation Learning for Point Cloud Analysis (AAAI 2025 · 2 citations) |
ShapeNetPart · instance mIoU (higher is better)
9 papers, 2019–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 86.2 | Relation-Shape Convolutional Neural Network for Point Cloud Analysis (CVPR 2019 · 1,003 citations) |
| 2023 | 86.7 | PointClustering: Unsupervised Point Cloud Pre-Training Using Transformation Invariance in Clustering (CVPR 2023 · 20 citations) |
| 2024 | 87.2 | Interpretable3D: An Ad-Hoc Interpretable Classifier for 3D Point Clouds (AAAI 2024 · 23 citations) |
ModelNet40 z/z · accuracy (higher is better)
8 papers, 2019–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 92.3 | Spherical Fractal Convolutional Neural Networks for Point Cloud Recognition (CVPR 2019 · 180 citations) |
ModelNet40 z/SO(3) · accuracy (higher is better)
7 papers, 2019–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 87.1 | ClusterNet: Deep Hierarchical Cluster Network With Rigorously Rotation-Invariant Representation for Point Cloud Analysis (CVPR 2019 · 193 citations) |
| 2020 | 91 | Rotation-Invariant Local-to-Global Representation Learning for 3D Point Cloud (NeurIPS 2020 · 46 citations) |
| 2023 | 91.8 | CRIN: Rotation-Invariant Point Cloud Analysis and Rotation Estimation via Centrifugal Reference Frame (AAAI 2023 · 9 citations) |
What has been tried
132 papers here report 1,423 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 5%, 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 |
|---|---|---|---|---|
| use all four serialization patterns instead of Z-order only | ScanNet semantic segmentation val · mIoU | 74.3 -> 76.5 | Helped | Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · 600 citations) |
| Add depthwise convolution | S3DIS Area 5 · mIoU | 68.5 ± 0.2 -> 68.0 ± 0.3 | Hurt | KPConvX: Modernizing Kernel Point Convolution with Kernel Attention (CVPR 2024 · 28 citations) |
| Add MLGC | PUNet 10K, 1% noise · CD | 25.93 -> 18.25 | Helped | Denoising Point Clouds in Latent Space via Graph Convolution and Invertible Neural Network (CVPR 2024 · 26 citations) |
| Add Sinkhorn-Knopp Prototype Estimation to random prototype initialization | ScanObjectNN · OA (%) | 68.0±1.5 -> 80.3±0.5 | Helped | Interpretable3D: An Ad-Hoc Interpretable Classifier for 3D Point Clouds (AAAI 2024 · 23 citations) |
| Use a shared structure kernel instead of concatenating explicit structure | ScanObjectNN · OA | 88.3 -> 88.8 | Helped | X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · 19 citations) |
| Directly regress parameterized edges rather than edge endpoints | Building3D Entry-level test · edge recall (ER) | 0.12 -> 0.82 | Helped | PBWR: Parametric-Building-Wireframe Reconstruction from Aerial LiDAR Point Clouds (CVPR 2024 · 17 citations) |
| Use Hop Graph Attention instead of SA with λ=0 and no hop distance loss | ModelNet40 · accuracy | +0.5% | Helped | DHGCN: Dynamic Hop Graph Convolution Network for Self-Supervised Point Cloud Learning (AAAI 2024 · 15 citations) |
| add size-aware encoding layer | SUSTech1K overall · Rank-1 accuracy | 77.1% to 80.8% | Helped | LidarGait++: Learning Local Features and Size Awareness from LiDAR Point Clouds for 3D Gait Recognition (CVPR 2025 · 9 citations) |
| Add PDAM to PLAM | S3DIS Area-5 · mIoU (%) | 69.5±0.3 -> 70.4±0.3 | Helped | Point Deformable Network with Enhanced Normal Embedding for Point Cloud Analysis (AAAI 2024 · 7 citations) |
| increase the number of steerable neurons beyond K=2 | ScanObjectNN OBJ_BG · accuracy (%) | 87.3; 84.5; 86.2; 85.4 | Mixed | TetraSphere: A Neural Descriptor for O(3)-Invariant Point Cloud Analysis (CVPR 2024 · 7 citations) |
| Increase network depth from 24 to 120 blocks | S3DIS Area5 · mIoU | 73.2 -> 75.7 | Helped | DeepLA-Net: Very Deep Local Aggregation Networks for Point Cloud Analysis (CVPR 2025 · 5 citations) |
| Use DP points instead of equidistant points | nuScenes val · mAP | 51.6 -> 52.2 | Helped | Compact HD Map Construction via Douglas-Peucker Point Transformer (AAAI 2024 · 5 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Baseline | S3DIS Area 5 | mIoU | 55.88 – 70.9 | 3 papers, e.g. CVPR 2019 and CVPR 2024 |
| PointNet++ | S3DIS | mIoU (%) | 47.6 – 56.36 | 2 papers, e.g. NeurIPS 2018 and AAAI 2021 |
| PointNet | S3DIS 6-fold validation | mIOU | 47.71 – 53.2 | 2 papers, e.g. CVPR 2018 and CVPR 2020 |
What reviewers will expect
Datasets and baselines named in the 16 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 · 6 of 16 papers
- ScanObjectNN · 5 of 16 papers
- ShapeNetPart · 3 of 16 papers
- S3DIS · 3 of 16 papers
- ScanObjectNN-C · 2 of 16 papers
- ScanNet · 2 of 16 papers
Baselines
- PointNet++ · 4 of 16 papers
- PointMLP · 3 of 16 papers
- DGCNN · 3 of 16 papers
- PointTransformer · 3 of 16 papers
- PointNeXt · 2 of 16 papers
- PTv3 · 2 of 16 papers
- MATE · 2 of 16 papers
- BFTT3D · 2 of 16 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 73 papers; from OpenAlex and the papers' affiliations).
- Peking University · 5 papers
- Tencent · 3 papers
- The University of Sydney · 3 papers
- Shanghai Jiao Tong University · 2 papers
- Xi’an Jiaotong University · 2 papers
- Google · 2 papers
- Seoul National University · 2 papers
- Nanjing University of Science and Technology · 2 papers
- Tsinghua University · 2 papers
- University of Michigan · 2 papers
Use it
Download BibTeX of 39 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
- Equivariant point cloud networks (84 papers)
- Efficient point sampling architectures (45 papers)
- Adversarial point cloud attacks (37 papers)
Elsewhere in the topic
- Point-language and open-vocabulary 3D (107 papers)
- Scene flow and dynamic point clouds (89 papers)
- Self-supervised point cloud pretraining (73 papers)
- LiDAR point cloud compression and completion (60 papers)
- Point cloud detection and tracking (56 papers)
- Masked point autoencoders (22 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).