Research map / Point clouds / Point cloud classification
LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling
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 LTA-PCS, a label-free, task-agnostic learnable point cloud sampler trained with geometric loss plus a semantic loss that compares frozen CLIP vision features of multi-view depth projections of dense and sampled clouds.
Problem
Point cloud sampling for downstream tasks when the task is unknown: FPS ignores semantics, while learnable sampling methods need task-specific labels or losses and generalize poorly.
Approach
Train a point sampling network without task labels by using a frozen pretrained 2D CLIP vision encoder on multi-view depth projections, with intra-view (per-view feature MSE) and inter-view (feature Gram/relation MSE) semantic losses added to the S-Net geometric loss; an inference-time nearest-neighbor matching makes the output a subset of the input.
Key results
- On ModelNet40 with 256 sampled points, LTA-PCS reaches 90.0% classification accuracy versus 87.9% for SampleNet* (task-agnostic), a 2.1% gain.
- On point cloud retrieval with 512 points on ModelNet40, LTA-PCS achieves 77.2 mAP versus 74.7 for SampleNet*.
- Ablation shows removing both semantic losses drops 256-point accuracy from 90.0 to 88.2 on ModelNet40.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LTA-PCS instead of the strongest listed baseline | PointNet classification · ModelNet40, 256 points | instance accuracy: 88.2 -> 90.0 | Helped |
| use LTA-PCS instead of the strongest listed baseline | PointNet classification · ShapeNet, 512 points | instance accuracy: 83.6 -> 84.4 | Helped |
| use LTA-PCS instead of the strongest listed baseline | PointNet retrieval · ModelNet40, 512 points | mAP: 74.7 -> 77.2 | Helped |
| add both intra-view and inter-view semantic losses (ablation) | LTA-PCS · ModelNet40, 128 points | instance accuracy: 87.8 -> 89.1 | Helped |
| add inter-view loss (ablation) | LTA-PCS · ModelNet40, 128 points | instance accuracy: 88.6 -> 89.1 | Helped |
| add intra-view loss (ablation) | LTA-PCS · ModelNet40, 128 points | instance accuracy: 88.3 -> 89.1 | Helped |
| replace PointNet with a CLIP-pretrained 2D feature extractor (ablation) | LTA-PCS · ModelNet40, 128 points | instance accuracy: 88.0 -> 89.1 | Helped |
| replace ImageNet-trained ResNet-50 with CLIP vision encoder ResNet-50 (ablation) | LTA-PCS · ModelNet40, 128 points | instance accuracy: 88.3 -> 89.1 | Helped |
| increase projected views from 2 to 10 (ablation) | LTA-PCS · ModelNet40, 128 points | instance accuracy: 88.2 -> 89.1 | Helped |
Surprising findings
- Using a 3D PointNet feature extractor for the semantic loss performs worse than the geometric-only setting, with 88.0 vs 89.1 at 128 points.
- LTA-PCS outperforms task-specific S-Net and SampleNet despite not using task labels.
Limitations
- Semantic supervision depends on projecting 3D points to depth maps, which loses information.
- Relies on a 2D CLIP encoder; no pretrained 3D feature extractor of comparable quality is available.
- Evaluation covers only classification, retrieval and completion on small shape datasets.
Experimental setup
Datasets: ModelNet40 ShapeNet Core55 ScanObjectNN
Benchmarks: ModelNet40 ShapeNet ScanObjectNN ShapeNet Core55
Metrics: instance accuracy mAP L1 Chamfer Distance
Baselines: FPS Random S-Net SampleNet S-Net* SampleNet* PCN
Builds on: S-Net SampleNet CLIP PointNet
Closest papers by meaning
- MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing (CVPR 2025 · similarity 0.89)
- GCD-Sampling: A General Cross-scale Decoupled Sampling for Point Cloud (AAAI 2025 · similarity 0.88)
- SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity (CVPR 2025 · similarity 0.86)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.85)
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.85)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.85)
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.