Research map / Point clouds / Point cloud classification
Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation
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 SinPoint, a sine-displacement point cloud augmentation (single or multiple sine functions) argued to be a homeomorphism, combined with a Markov-chain augmentation process, improving point cloud classification and part segmentation.
Problem
Existing point cloud data augmentation methods either fail to enrich geometric diversity or break topological consistency (e.g., mixing or non-rigid deformation can distort shapes), harming fine-grained point cloud understanding.
Approach
Treat augmentation as a residual displacement P' = P + A sin(wP + phi) that is argued to be a homeomorphism when -1 <= Aw <= 1, so topology and labels are preserved; superpose multiple sine fields anchored at points (SinPoint-MSF) and chain basic transformations via a Markov process.
Key results
- SinPoint with DGCNN reaches 93.7 OA on ModelNet40 (+2.0 over baseline 91.7 in ablation) and 84.6 on PB T50 RS (+7.3 over 77.3).
- With PointNet++ on OBJ ONLY, SinPoint reaches 90.2 OA (+5.9 over 84.3), matching the paper's claimed state-of-the-art real-world accuracy.
- On ShapeNetPart, SinPoint-MSF improves mIoU by 1.0 for PointNet++ (84.8 to 85.8) and 0.7 for DGCNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SinPoint-SSF | PointNet · OBJ ONLY | overall accuracy: 76.1 -> 82.6 | Helped |
| Apply SinPoint-SSF | PointNet++ · OBJ ONLY | overall accuracy: 84.3 -> 90.2 | Helped |
| Apply SinPoint-SSF | DGCNN · PB T50 RS | overall accuracy: 77.3 -> 84.6 | Helped |
| Add SinPoint (ablation) | DGCNN · ModelNet40 | overall accuracy: 91.7 -> 92.7 | Helped |
| Add Markov augmentation process (ablation) | DGCNN · ModelNet40 | overall accuracy: 91.7 -> 92.9 | Helped |
| Add Markov augmentation process to SinPoint (ablation) | DGCNN · ModelNet40 | overall accuracy: 92.7 -> 93.0 | Helped |
| Mix original and augmented samples (ablation) | DGCNN with SinPoint · ModelNet40 | overall accuracy: 92.7 -> 93.4 | Helped |
| Add sample mixing to SinPoint and Markov augmentation (ablation) | DGCNN · ModelNet40 | overall accuracy: 93.0 -> 93.7 | Helped |
| Use SinPoint-MSF instead of SinPoint-SSF (ablation) | DGCNN · OBJ ONLY | overall accuracy: 90.2 -> 89.8 | Hurt |
| Use SinPoint-MSF instead of SinPoint-SSF (ablation) | DGCNN · ShapeNetPart | mIoU: 85.3 -> 85.5 | Helped |
| Use uniform rather than Gaussian sampling for amplitude and angular velocity (ablation) | SinPoint with DGCNN · ModelNet40 | overall accuracy: 87.607±0.344 -> 89.759±0.431 | Helped |
Surprising findings
- The paper reports that DGCNN with SinPoint is only 0.4% below PointMLP while using about 7x fewer parameters.
Limitations
- The homeomorphism claim holds only under the condition -1 <= A*omega <= 1; otherwise the mapping is only a local homeomorphism, and sampled ranges may violate it.
- Theorem 1 (variance increase) is argued from a loose definition, and the paper provides no hardware or training cost details.
Experimental setup
Datasets: ModelNet40 ReducedMN40 ScanObjectNN (OBJ ONLY) ScanObjectNN (PB T50 RS) ShapeNetPart S3DIS SemanticKITTI
Benchmarks: ModelNet40 ReducedMN40 ScanObjectNN OBJ ONLY ScanObjectNN PB T50 RS ShapeNetPart S3DIS SemanticKITTI
Metrics: overall accuracy mIoU mAcc
Baselines: PointAugment PointMixup PatchAugment PointWOLF RSMix SageMix PointCutMix WOLFMix PCSalMix PointPatchMix
Builds on: PointWOLF PointMixup SageMix ResNet residual connection
Closest papers by meaning
- SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (CVPR 2024 · similarity 0.87)
- SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds (ICML 2025 · similarity 0.86)
- An Explicit Frame Construction for Normalizing 3D Point Clouds (ICML 2024 · similarity 0.86)
- Point Deformable Network with Enhanced Normal Embedding for Point Cloud Analysis (AAAI 2024 · similarity 0.86)
- Point-Level Topological Representation Learning on Point Clouds (ICML 2025 · similarity 0.86)
- PointPatchMix: Point Cloud Mixing with Patch Scoring (AAAI 2024 · similarity 0.86)
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.