Research map / Point clouds / Point cloud classification
GCD-Sampling: A General Cross-scale Decoupled Sampling for Point Cloud
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 GCD-sampling, a coordinate-only, parameter-shared adaptive point cloud sampling layer that learns convex combinations of k-NN points around FPS seeds, embeddable in hierarchical point cloud networks.
Problem
Fixed sampling such as farthest point sampling is suboptimal for downstream tasks, while existing adaptive point cloud sampling methods are over-coupled with specific target networks and lack multi-scale reuse, making them model-specific.
Approach
Decouple sampling from the target network by using only point coordinates as input and output, and share one sampling module across all layers via an arbitrary-scale design with a cached global feature and convex combination learning of k-NN neighbors of FPS points.
Key results
- On ModelNet40 with PointNeXt, GCD-sampling reaches 93.35% OA vs 93.15% for FPS, adding only 0.08M parameters.
- On ScanObjectNN with PointNeXt, GCD-sampling reaches 87.86% OA vs 87.40% for FPS and 87.65% for APES.
- On ShapeNetPart with PointNet++, GCD-sampling reaches 85.54% ins.mIoU vs 85.06% for FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FPS with GCD-sampling | PointNet++, PointCNN, PointMLP, and PointNeXt · ModelNet40 and ScanObjectNN (PB T50 RS) | OA (%): 92.06 -> 92.10; 78.07 -> 78.35; 92.38 -> 92.59; 78.90 -> 79.11; 93.11 -> 93.31; 87.34 -> 87.54; 93.27 -> 93.35; 87.65 -> | Helped |
| Replace FPS with GCD-sampling | PointNet++, PointCNN, PointMLP, and PointNeXt · ShapeNetPart | ins.mIoU (%): 85.25 -> 85.54; 86.31 -> 86.50; 86.18 -> 86.45; 86.78 -> 86.90 | Helped |
| Fine-tune only GCD-sampling after loading pretrained PointNeXt weights (ablation) | PointNeXt · ScanObjectNN (PB T50 RS) | OA (%): 87.40 -> 88.41 | Helped |
| Load pretrained PointNeXt weights and unfreeze all parameters instead of training only GCD-sampling (ablation) | PointNeXt + GCD-sampling · ScanObjectNN (PB T50 RS) | OA (%): 88.41 -> 87.86 | Hurt |
| Use GenDirect instead of convex-combination learning (ablation) | GCD-sampling with PointNet++ and PointCNN · ScanObjectNN (PB T50 RS) | OA (%): 78.35 -> 53.73; 79.11 -> 54.17 | Hurt |
| Use GenRes instead of convex-combination learning (ablation) | GCD-sampling with PointNet++ and PointCNN · ScanObjectNN (PB T50 RS) | OA (%): 78.35 -> 58.67; 79.11 -> 59.65 | Hurt |
| Add Chamfer distance loss to GenDirect (ablation) | GCD-sampling with PointNet++ and PointCNN · ScanObjectNN (PB T50 RS) | OA (%): 53.73 -> 75.67; 54.17 -> 76.97 | Helped |
| Add Chamfer distance loss to GenRes (ablation) | GCD-sampling with PointNet++ and PointCNN · ScanObjectNN (PB T50 RS) | OA (%): 58.67 -> 78.07; 59.65 -> 78.78 | Helped |
| Add Chamfer distance loss to GenConvex (ablation) | GCD-sampling with PointNet++ · ScanObjectNN (PB T50 RS) | OA (%): 78.35 -> 78.10 | Hurt |
| Increase the convex-learning k-nearest-neighbor count from 1 toward 10 (ablation) | GCD-sampling · ScanObjectNN (PB T50 RS) | OA (%) | Helped |
| Increase k excessively beyond the effective range (ablation) | GCD-sampling · ScanObjectNN (PB T50 RS) | OA (%) | Hurt |
Surprising findings
- Fine-tuning only the 0.08M GCD-sampling parameters on a frozen pretrained PointNeXt raised ScanObjectNN OA to 88.41%, above training the whole network from scratch.
- Adding a chamfer distance loss to the convex-combination variant slightly reduced performance (e.g., 87.54 to 87.25 on PointNeXt ModelNet40 setting).
Limitations
- Gains are modest (often under 1 percentage point) relative to FPS and baselines.
- Only classification and part segmentation are evaluated; no detection or semantic segmentation.
Experimental setup
Datasets: ModelNet40 ScanObjectNN ShapeNetPart
Benchmarks: ModelNet40 ScanObjectNN ShapeNetPart
Metrics: overall accuracy mean class accuracy instance mIoU class mIoU parameters FLOPS inference time
Baselines: FPS ASNL-AS APES CPL GSS
Builds on: PointNet++ FPS ASNL APES PointNeXt
Closest papers by meaning
- SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity (CVPR 2025 · similarity 0.89)
- LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling (CVPR 2024 · similarity 0.88)
- MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing (CVPR 2025 · similarity 0.87)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · similarity 0.86)
- CRA-PCN: Point Cloud Completion with Intra- and Inter-level Cross-Resolution Transformers (AAAI 2024 · similarity 0.86)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · 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.