Research map / Point clouds / Point cloud classification
LinNet: Linear Network for Efficient Point Cloud Representation Learning
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.
Proposes LinNet, a purely point-based MLP network with a disassembled set abstraction module and linear-complexity space-filling-curve hash query and linearization sampling, achieving state-of-the-art large-scale segmentation with ~10x speedup over PointNeXt.
Problem
Point-based point cloud networks do not scale to large-scale 3D scenes because FPS and KNN neighborhood search have high time complexity and set abstraction is computationally expensive.
Approach
Decouple set abstraction into separate point-wise semantic transform and neighborhood geometric bias (DSA, inspired by depthwise-separable convolutions), and replace FPS/KNN with space-filling-curve-based shuffled-key hash querying and linearization sampling, giving linear complexity on GPUs.
Key results
- LinNet achieves 73.7% mIoU on S3DIS Area 5 with TTA (72.9% without), outperforming PointNeXt's 70.8%.
- LinNet reaches 81.4% mIoU on NuScenes validation with TTA and 69.1% on SemanticKITTI validation.
- At 200k points LinNet runs about 13x faster than PointNeXt; ablation shows latency reduced from 89 ms to 34 ms while mIoU rises from 70.8 to 73.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LinNet with test-time augmentation instead of the listed strongest baseline | S3DIS Area 5 semantic segmentation · S3DIS Area 5 | mIoU: 73.5 -> 73.7 | Mixed |
| Use LinNet with test-time augmentation instead of PTv3 with TTA | NuScenes semantic segmentation · NuScenes val | mIoU: 80.4 -> 81.4 | Helped |
| Use LinNet instead of 2DPASS | SemanticKITTI semantic segmentation · SemanticKITTI val | mIoU: 69.3 -> 69.1 | Hurt |
| Use LinNet instead of PointMetaBase | ScanObjectNN PB_T50_RS classification · ScanObjectNN PB_T50_RS | OA: 87.9±0.2 -> 88.2±0.4 | Helped |
| Use LinNet instead of PointMLP | ModelNet40 classification · ModelNet40 | OA: 94.1 -> 93.6±0.2 | Hurt |
| Add linearization sampling to the PointNeXt baseline (ablation) | S3DIS Area 5 model design · S3DIS Area 5, 20k points | mIoU: 70.8 -> 72.0 | Helped |
| Add hash query to the PointNeXt baseline (ablation) | S3DIS Area 5 model design · S3DIS Area 5, 20k points | mIoU: 70.8 -> 70.8 | No clear effect |
| Add linearization sampling and hash query to the PointNeXt baseline (ablation) | S3DIS Area 5 model design · S3DIS Area 5, 20k points | latency (ms): 89 -> 38 | Helped |
| Replace SA with depth-wise separate set abstraction (ablation) | S3DIS Area 5 model design · S3DIS Area 5, 20k points | mIoU: 72.0 -> 71.0 | Hurt |
| Replace DSSA with DSA (ablation) | S3DIS Area 5 model design · S3DIS Area 5, 20k points | mIoU: 71.0 -> 73.1 | Helped |
| Use DSA instead of vanilla SA (ablation) | S3DIS Area 5 feature aggregation · S3DIS Area 5 | mIoU: 72.0 -> 73.1 | Helped |
| Use DSA instead of ASSA (ablation) | S3DIS Area 5 feature aggregation · S3DIS Area 5 | mIoU: 70.5 -> 73.1 | Helped |
Surprising findings
- Plain separation of semantic and geometric inputs (DSSA) lowers accuracy by about 1%, and the separated-input form degrades later in training despite faster early convergence.
- LinNet-Small with 1.5-1.7M parameters surpasses the 38M-parameter MinkUNet on NuScenes validation (77.6% vs 73.3%).
Limitations
- Point-wise neighborhood aggregation and discontinuous memory layout lead to high memory footprint and cache misses, higher than sparse convolution methods.
- Memory usage is higher than MinkUNet (16 GB training for LinNet vs 2.6 GB for MinkUNet).
Experimental setup
Datasets: S3DIS NuScenes SemanticKITTI ScanObjectNN ModelNet40
Benchmarks: S3DIS Area 5 NuScenes SemanticKITTI ScanObjectNN ModelNet40
Metrics: mIoU OA mAcc latency throughput FLOPs parameters
Baselines: PointNeXt Point Transformer v2 Stratified Transformer PointVector PointMetaBase MinkUNet SphereFormer PTv3
Builds on: PointNet++ PointNeXt MobileNet Point Transformer V3
Closest papers by meaning
- ASSANet: An Anisotropic Separable Set Abstraction for Efficient Point Cloud Representation Learning (NeurIPS 2021 · similarity 0.92)
- PointCNN++: Performant Convolution on Native Points (CVPR 2026 · similarity 0.90)
- PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation (CVPR 2017 · similarity 0.89)
- Mining Point Cloud Local Structures by Kernel Correlation and Graph Pooling (CVPR 2018 · similarity 0.89)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.89)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.88)
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.