Research map / Point clouds / Point cloud classification
ViewNet: A Novel Projection-Based Backbone With View Pooling for Few-Shot Point Cloud Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace DGCNN backbone with ViewNet (projection-based) while keeping CIA head | few-shot point cloud classification · ScanObjectNN 5-way 1-shot | accuracy: +5.38 points over CIA | Helped |
| Replace DGCNN backbone with ViewNet while keeping CIA head | few-shot point cloud classification · ModelNet40-C 5-way 1-shot | accuracy: +2.75 points on mean over CIA | Helped |
| Replace DGCNN backbone with ViewNet while keeping CIA head | few-shot point cloud classification · ModelNet40 5-way 1-shot | accuracy: +1.96 points over CIA | Helped |
| Replace DGCNN backbone with ViewNet while keeping CIA head | few-shot point cloud classification · ScanObjectNN 5-way 5-shot | accuracy: +6.17 points over ProtoNet (second best) | Helped |
| Replace DGCNN backbone with ViewNet while keeping CIA head | few-shot point cloud classification · ModelNet40 5-way 5-shot | accuracy: +1.58 points over CIA | Helped |
| Compress ViewNet bin-wise feature matrix O into a single vector O' before the few-shot head (remove bin-wise loss) (ablation) | ViewNet+CIA · ScanObjectNN 5-way 1-shot | accuracy: 63.83% -> 61.43% mean | Hurt |
| Compress ViewNet bin-wise feature matrix O into a single vector O' before the few-shot head (ablation) | ViewNet+CIA · ScanObjectNN 5-way 5-shot | accuracy: 75.30% -> 74.37% mean | Hurt |
| Add View Pooling combinations (opposite-view pairs and triplets, Ci) to the pooled features (ablation) | ViewNet · ScanObjectNN 5-way 1-shot | accuracy: 62.73% -> 63.83% mean | Helped |
| Add View Pooling combinations (opposite-view pairs and triplets, Ci) to the pooled features (ablation) | ViewNet · ScanObjectNN 5-way 5-shot | accuracy: 73.92% -> 75.30% mean | Helped |
| Replace DGCNN with ViewNet as backbone (with compressed output) across four few-shot heads (ablation) | MetaOpt, RelationNet, ProtoNet, CIA heads · ScanObjectNN 5-way 1-shot and 5-shot | accuracy: Mean gains of 4.14 to 6.39 points in 1-shot and 4.78 to 6.41 points in 5-shot, except RelationNet 5-shot head in fold 1 | Helped |
| Replace DGCNN with ViewNet as backbone under ProtoNet head on ModelNet40 (ResNet vs DGCNN comparison also shown) | ResNet backbone with ProtoNet vs DGCNN backbone · ModelNet40 5-way 1-shot | accuracy: ResNet+ProtoNet 75.38% vs DGCNN+ProtoNet 76.42% mean | Hurt |
Closest papers by meaning
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.89)
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.88)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.88)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.88)
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.88)
- LPCG: A Self-conditional Architecture for Labeled Point Cloud Generation (AAAI 2025 · similarity 0.88)
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.