Research map / Point clouds / Point cloud classification
PointCLIP: Point Cloud Understanding by CLIP
Check 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 |
|---|---|---|---|
| Project point clouds from multiple views and classify with frozen CLIP | PointCLIP · ModelNet10, ModelNet40, and ScanObjectNN | accuracy | Helped |
| Fine-tune the inter-view adapter with 16 examples per category | PointCLIP with CLIP ResNet-101 · ModelNet40 | accuracy: from 20.18\% to 87.20\% | Helped |
| Ensemble 16-shot PointCLIP logits with classical 3D network logits | PointNet++, RSCNN, DGCNN, SimpleView, and CurveNet · ModelNet40 | accuracy: PointNet++: 89.71 to 92.10; CurveNet: 93.84 to 94.08 | Helped |
| Increase zero-shot projections from 6 to more than 6 views (ablation) | PointCLIP · ModelNet40, zero-shot | accuracy: 20.18 -> 16.98; 14.91; 13.65 | Hurt |
| Increase few-shot projections from 6 to 10 views (ablation) | PointCLIP with inter-view adapter · ModelNet40, 16-shot | accuracy: 84.24 -> 87.20 | Helped |
| Remove global feature extraction from the inter-view adapter (ablation) | PointCLIP with inter-view adapter · ModelNet40, 16-shot | accuracy: 87.20 -> 83.87 | Hurt |
| Remove view-wise adapted feature generation (ablation) | PointCLIP with inter-view adapter · ModelNet40, 16-shot | accuracy: 87.20 -> 85.93 | Hurt |
| Reduce the inter-view adapter to two layers (ablation) | PointCLIP with inter-view adapter · ModelNet40, 16-shot | accuracy: 87.20 -> 86.48 | Hurt |
| Move a linear layer before global feature extraction (ablation) | PointCLIP with inter-view adapter · ModelNet40, 16-shot | accuracy: 87.20 -> 86.78 | Hurt |
| Replace the depth-map prompt with the best few-shot prompt, “point cloud of a big [CLASS].” (ablation) | PointCLIP · ModelNet40 | accuracy: zero-shot: 20.18\% -> 19.21\%; 16-shot: 85.82\% -> 87.20\% | Mixed |
| Use 16-shot PointCLIP rather than zero-shot PointCLIP in the CurveNet ensemble (ablation) | CurveNet + PointCLIP · ModelNet40 | accuracy: 93.88 -> 94.08 -> 93.88 | Mixed |
| Ensemble two models with the same training scheme instead of combining PointCLIP with a 3D-trained model (ablation) | ModelNet40 model ensembles · ModelNet40 | accuracy: PointNet++ + CurveNet: 91.61; SimpleView + CurveNet: 93.68; PointCLIP + PointCLIP: 87.06 | Hurt |
Closest papers by meaning
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.90)
- Frozen CLIP Transformer Is an Efficient Point Cloud Encoder (AAAI 2024 · similarity 0.90)
- Duoduo CLIP: Efficient 3D Understanding with Multi-View Images (ICLR 2025 · similarity 0.89)
- Occlusion-aware Text-Image-Point Cloud Pretraining for Open-World 3D Object Recognition (CVPR 2025 · similarity 0.88)
- SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language (CVPR 2025 · similarity 0.88)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · 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.