Research map / Point clouds / Shape and surface completion
SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity
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 SAMBLE, a learning-to-sample method for point clouds that uses a sparse attention map and adaptive bin-based sampling to learn shape-specific sampling strategies balancing local detail and global uniformity.
Problem
Learning-based point cloud sampling either generates new points (losing interpretability) or over-focuses on sharp edge points, producing biased samples and poor few-point sampling; existing methods also apply one sampling strategy to all shapes.
Approach
A sparse attention map built by carving k-NN cells from a global attention map, which yields point-wise scores via column-based indexing (column square-divided), plus adaptive momentum-updated bin boundaries and learned bin tokens that set per-shape bin sampling ratios, letting sampling go beyond top-M to include close-to-edge and non-edge points.
Key results
- Classification OA 94.2% on ModelNet40, above APES (global) 93.8% and PointNeXt 93.2%.
- Part segmentation reaches 84.5% category mIoU and 86.7% instance mIoU on ShapeNet-Part, state of the art among compared methods.
- Few-point sampling on ModelNet40 with PointNet at M=32 reaches 89.45% vs. APES (w/ pre-processing) 88.56%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAMBLE instead of CurveNet | ModelNet40 classification · ModelNet40 | OA (%): 93.8 -> 94.2 | Helped |
| Use SAMBLE instead of PointNeXt | ShapeNet-Part segmentation · ShapeNet-Part | Cat. mIoU (%): 84.4 -> 84.5 | Helped |
| Use SAMBLE instead of PointNeXt | ShapeNet-Part segmentation · ShapeNet-Part | Ins. mIoU (%): 86.7 -> 86.7 | No clear effect |
| Use SAMBLE sampling rather than PointNeXt/FPS sampling | ShapeNet-Part segmentation at intermediate point counts · ShapeNet-Part, 1024 and 512 points | Cat. mIoU (%): 83.79 -> 84.84; 82.77 -> 85.04 | Helped |
| Use SAMBLE sampling rather than PointNeXt/FPS sampling | ShapeNet-Part segmentation at intermediate point counts · ShapeNet-Part, 1024 and 512 points | Ins. mIoU (%): 86.18 -> 86.93; 85.18 -> 87.12 | Helped |
| Use SAMBLE instead of APES with pre-processing | ModelNet40 few-point classification · ModelNet40, 128, 64, and 32 points | accuracy: 89.77 -> 90.02; 89.57 -> 89.81; 88.56 -> 89.45 | Helped |
| Use sparse column square-divided indexing (mode vii) rather than sparse column sum (mode v) (ablation) | SAMBLE classification · ModelNet40 | OA (%): 93.40 -> 94.08 | Helped |
| Use sparse column square-divided indexing (mode vii) rather than sparse column sum (mode v) (ablation) | SAMBLE segmentation · ShapeNet-Part | Cat. mIoU (%): 83.47 -> 84.22 | Helped |
| Increase the number of bins from 1 to 6 (ablation) | SAMBLE classification · ModelNet40 | OA (%): 93.95 -> 94.18 | Helped |
| Increase the number of bins from 1 to 4 (ablation) | SAMBLE segmentation · ShapeNet-Part | Cat. mIoU (%): 84.22 -> 84.51 | Helped |
| Replace cross-attention upsampling with interpolation (ablation) | SAMBLE segmentation · ShapeNet-Part | Cat. mIoU (%): 84.51 -> 84.35 | No clear effect |
| Replace cross-attention upsampling with interpolation (ablation) | APES (local) segmentation · ShapeNet-Part | Cat. mIoU (%): 82.89 -> 82.95 | Hurt |
Surprising findings
- Increasing the number of bins does not improve performance; the best settings were 6 bins for classification and 4 for segmentation.
- APES degrades when interpolation replaces its cross-attention upsampling, whereas SAMBLE improves with interpolation.
Limitations
- Upsampling layer is not optimized; the authors note further gains are possible with a better upsampling design.
- Evaluated only on shape-level datasets (ModelNet40, ShapeNet-Part); extension to point cloud scenes is left open.
- Performance is sensitive to the number of bins, with more bins degrading results.
Experimental setup
Datasets: ModelNet40 ShapeNet-Part
Benchmarks: ModelNet40 ShapeNet-Part
Metrics: overall accuracy category mIoU instance mIoU
Baselines: APES FPS RS S-Net SampleNet MOPS-Net PointNeXt LighTN PST-NET DA-Net
Builds on: APES PointNet++ PointNeXt Vision Transformer
Closest papers by meaning
- GCD-Sampling: A General Cross-scale Decoupled Sampling for Point Cloud (AAAI 2025 · similarity 0.89)
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.87)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.87)
- MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing (CVPR 2025 · similarity 0.87)
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.87)
- 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.