Research map / Point clouds / Segmentation and detection
Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding
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 an abstraction-and-refinement (ABS-REF) view of point cloud models and a High-dimensional Positional Encoding (HPE) module, used to build MLP-based HPENets that reach SOTA on ScanObjectNN, ShapeNetPart and S3DIS.
Problem
MLP-based point cloud processing lacks a clear understanding of what drives performance, and existing positional encodings are low-dimensional or non-adaptive, under-using positional information.
Approach
Views point cloud models as ABS (sampling/grouping/aggregation) plus REF (inter-set refinement) stages, and adds a high-dimensional positional encoding (sinusoidal or MLP-learned) in both stages of an MLP network, lifting 3D coordinates to a high-dimensional space before aligning them to feature channels.
Key results
- HPENet(MLP) reaches 87.6% mAcc and 88.9% OA on ScanObjectNN PB_T50_RS, the SOTA among methods without pre-training, voting or ensembling.
- HPENet achieves 72.7% mIoU on S3DIS Area-5 and 78.7% mIoU on S3DIS 6-fold, and 85.5 class mIoU on ShapeNetPart.
- GroupFree3D + HPE(MLP) improves ScanNet V2 detection to 69.1 mAP@0.25 and 53.0 mAP@0.5 versus the GroupFree3D baseline of 68.2 and 52.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HPE(MLP) to the VoteNet detection model | VoteNet · ScanNet V2 | mAP@0.25: 63.8 -> 65.0 | Helped |
| Add HPE(MLP) to the VoteNet detection model | VoteNet · ScanNet V2 | mAP@0.5: 44.2 -> 45.6 | Helped |
| Add HPE(MLP) to the GroupFree3D detection model | GroupFree3D · ScanNet V2 | mAP@0.25: 68.2 -> 69.1 | Helped |
| Add HPE(MLP) to the GroupFree3D detection model | GroupFree3D · ScanNet V2 | mAP@0.5: 52.6 -> 53.0 | Helped |
| Add HPESIN positional encoding to the degenerated HPENet (ablation) | HPENet-dv(32,[0,0,0,0]) · S3DIS Area-5 | mIoU: +0.8 | Helped |
| Add HPEMLP positional encoding to the degenerated HPENet (ablation) | HPENet-dv(32,[0,0,0,0]) · S3DIS Area-5 | mIoU: +1.1 | Helped |
| Add one REF layer to each ABS-REF block (ablation) | HPENet-dv(32,[0,0,0,0]) · S3DIS Area-5 | mIoU: +2.5 | Helped |
| Use HPEMLP and add one REF layer to each ABS-REF block (ablation) | HPENet-dv(32,[0,0,0,0]) · S3DIS Area-5 | mIoU: 64.2 -> 69.9 | Helped |
| Remove the grouping operation from the REF stage (ablation) | HPENet-dv(32,[1,1,1,1]) · S3DIS Area-5 | mIoU: 66.7 -> 63.7 | Hurt |
| Use HPEMLP in both ABS and REF stages rather than only one stage (ablation) | HPENet · S3DIS Area-5 | mIoU: 70.8 -> 72.7 | Helped |
| Use absolute instead of relative coordinates for HPESIN (ablation) | HPENet with HPESIN in ABS and REF · S3DIS Area-5 | mIoU: 72.4 -> 69.1 | Hurt |
| Replace element-wise addition with multiplication for HPESIN (ablation) | HPENet with HPESIN in ABS and REF · S3DIS Area-5 | mIoU: 72.4 -> 70.8 | Hurt |
Surprising findings
- Removing the REF stage's inter-set grouping (HPENet-dv*) drops S3DIS Area-5 mIoU from 69.9 to 63.7 at the same parameter count, showing the REF stage matters beyond parameter count.
- Element-wise multiplication of the sinusoidal positional encoding performs worse than addition, and absolute coordinates (HPESIN(abs)) reduce mIoU below the no-encoding baseline.
Limitations
- Claimed SOTA holds only for methods without pre-training, voting or ensembling.
- The ABS-REF view is a conceptual framing and the HPE design choices are validated mainly by ablations on S3DIS Area-5.
Experimental setup
Datasets: ScanObjectNN ModelNet40 ShapeNetPart S3DIS ScanNet V2
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart S3DIS Area-5 S3DIS 6-fold ScanNet V2
Metrics: mAcc OA mIoU class mIoU instance mIoU mAP@0.25 mAP@0.5
Baselines: PointNet PointNet++ DGCNN PointMLP PointNeXt PointMetaBase Point Transformer Stratified Transformer KPConv GroupFree3D 3DETR
Builds on: PointNet++ PointNeXt PointMLP Point Transformer VoteNet
Closest papers by meaning
- PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies (NeurIPS 2022 · similarity 0.89)
- Point Deformable Network with Enhanced Normal Embedding for Point Cloud Analysis (AAAI 2024 · similarity 0.88)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.87)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.87)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.87)
- PosDiffNet: Positional Neural Diffusion for Point Cloud Registration in a Large Field of View with Perturbations (AAAI 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.