Research map / Point clouds / Point cloud classification
CLIP2: Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data
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 |
|---|---|---|---|
| Train on point clouds with joint language-point and image-point alignment (ablation) | CLIP² · SUN RGB-D | Avg. Top1 accuracy: 56.5 -> 61.3 | Helped |
| Train on point clouds with joint language-point and image-point alignment (ablation) | CLIP² · ScanNet / SUN RGB-D / ScanObjectNN | Avg. Top1 accuracy: 52.6 / 26.8 / 37.7 -> 61.3 / 28.8 / 39.4 | Helped |
| Use point clouds instead of projected depth maps (ablation) | CLIP² · SUN RGB-D, nuScenes, ScanObjectNN | Avg. Top1 accuracy: 38.0 / 9.3 / 31.8 -> 61.3 / 28.8 / 39.4 | Helped |
| Align depth maps to image space rather than language space (ablation) | CLIP² depth representation · SUN RGB-D, nuScenes, ScanObjectNN | Avg. Top1 accuracy: 38.0 / 9.3 / 31.8 -> 56.4 / 21.7 / 39.0 | Helped |
| Use image-point alignment rather than language-point alignment (ablation) | CLIP² point-cloud representation · SUN RGB-D | Avg. Top1 accuracy: 52.6 -> 56.5 | Helped |
| Use image-point alignment rather than language-point alignment (ablation) | CLIP² point-cloud representation · nuScenes | Avg. Top1 accuracy: 26.8 -> 24.4 | Hurt |
| Replace PointNet++ with PointNet (ablation) | CLIP² · ScanNet | Avg. Top1 accuracy: 38.5 -> 22.6 | Hurt |
| Replace PointNet++ with DGCNN (ablation) | CLIP² · ScanNet | Avg. Top1 accuracy: 38.5 -> 26.0 | Hurt |
| Ensemble the point-cloud representation with image features (ablation) | CLIP² · SUN RGB-D | Avg. Top1 accuracy: 61.3 -> 68.7 | Helped |
| Ensemble the point-cloud representation with image features (ablation) | CLIP² · nuScenes | Avg. Top1 accuracy: 28.8 -> 43.9 | Helped |
| Add depth representation to the point-cloud and image ensemble (ablation) | CLIP² · SUN RGB-D | Avg. Top1 accuracy: 68.7 -> 69.6 | Helped |
| Add depth representation to the point-cloud and image ensemble (ablation) | CLIP² · nuScenes | Avg. Top1 accuracy: 43.9 -> 42.3 | Hurt |
Closest papers by meaning
- Vision-Language Pre-training with Object Contrastive Learning for 3D Scene Understanding (AAAI 2024 · similarity 0.91)
- Sculpting Holistic 3D Representation in Contrastive Language-Image-3D Pre-training (CVPR 2024 · similarity 0.91)
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.90)
- More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding (AAAI 2025 · similarity 0.90)
- RegionPLC: Regional Point-Language Contrastive Learning for Open-World 3D Scene Understanding (CVPR 2024 · similarity 0.90)
- CLIP-Driven Open-Vocabulary 3D Scene Graph Generation via Cross-Modality Contrastive Learning (CVPR 2024 · similarity 0.89)
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.