Research map / Point clouds / Point cloud classification
DeepLA-Net: Very Deep Local Aggregation Networks for Point Cloud Analysis
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.
Builds very deep local aggregation point cloud networks (DeepLA-Net, up to 120 blocks) using a cheap residual local feature extraction block and a hybrid deep supervision loss, reaching SOTA on S3DIS, ScanNet, ScanObjectNN and ShapeNetPart.
Problem
Local aggregation networks for point clouds have saturated in performance, and their costly local representations make it hard to build very deep networks that are also hard to optimize.
Approach
Starting from PointNet++, it introduces a ResLFE block (stage-level cached position embedding, front-linear vector feature representation, and a modernized residual structure without the 4x FFN expansion) that cuts FLOPs ~10x, plus a hybrid deep supervision strategy (per-stage semantic cross-entropy and spatial-distribution MSE with decaying weights) to train networks of 24 to 120 blocks.
Key results
- DeepLA-120 reaches 75.7% mIoU on S3DIS Area5 and 79.8% on 6-fold, surpassing the 75% threshold for the first time.
- DeepLA-24 reaches 90.6% OA on ScanObjectNN, the first method above 90% OA, and 86.5% Cls. mIoU on ShapeNetPart.
- The ResLFE block cuts FLOPs about 10x (92.2G for plain-24 to 9.2G for DeepLA-24) while improving mIoU from 69.0% to 70.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase network depth from 24 to 120 blocks (ablation) | DeepLA-Net · S3DIS Area5 | mIoU: 73.2 -> 75.7 | Helped |
| Use stage-level position embedding shared within each stage (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 69.0% → 68.6% | Hurt |
| Move feature abstraction before grouping (front-linear) (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 68.6% → 67.4% | Hurt |
| Use vector features instead of grouped features (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 67.4% → 68.6% | Helped |
| Replace concatenation with addition when combining vector features and position embeddings (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 68.6% → 68.4% | Hurt |
| Add the modernization structure (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 68.4% → 70.5% | Helped |
| Remove the 4× inverted bottleneck (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 70.5% → 70.1% | Hurt |
| Add semantic probability supervision (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 70.1% → 72.2% | Helped |
| Add spatial distribution supervision (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 70.1% → 71.8% | Helped |
| Use the hybrid loss combining semantic, spatial, and prediction losses (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 70.1% → 73.2% | Helped |
| Use a 0.5× bottleneck instead of 1× (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 73.2 -> 72.5 | Hurt |
| Use a 2× bottleneck instead of 1× (ablation) | DeepLA-24 · S3DIS Area5 | mIoU: 73.2 -> 73.4 | Helped |
Surprising findings
- Simply deepening a PointNet++-style baseline to 24 blocks increases FLOPs about 9x, yet a cheaper block design makes deeper networks both more accurate and faster than plain deep versions.
- Removing the 4x inverted bottleneck from the modernized FFN halves FLOPs with only a 0.4% mIoU drop, showing wide expansion is unnecessary for local aggregation.
Limitations
- Deeper variants (DeepLA-60/120) were not evaluated on ScanObjectNN or ShapeNetPart because of limited dataset scale.
- ScanNet test results are reported only for DeepLA-120 because of the online submission policy, with no validation results for the test-set run.
- Hyperparameters such as K, channel width and the HDS loss weights were tuned on S3DIS Area5.
Experimental setup
Datasets: S3DIS ScanNet v2 ScanObjectNN ShapeNetPart
Benchmarks: S3DIS Area5 S3DIS 6-fold ScanNet v2 val ScanNet v2 test ScanObjectNN ShapeNetPart
Metrics: mIoU OA mAcc Cls. mIoU Ins. mIoU FLOPs latency
Baselines: PointNeXt PointVector Point Transformer PointMeta KPConvX PDNet Point Transformer v3 OPFR
Builds on: PointNet++ PointNeXt Transformer ResNet
Closest papers by meaning
- RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds (CVPR 2020 · similarity 0.90)
- PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies (NeurIPS 2022 · similarity 0.89)
- PointCNN++: Performant Convolution on Native Points (CVPR 2026 · similarity 0.89)
- PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation (CVPR 2017 · similarity 0.89)
- SO-Net: Self-Organizing Network for Point Cloud Analysis (CVPR 2018 · similarity 0.88)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (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.