Research map / Point clouds / Point cloud classification
CLIP2Point: Transfer CLIP to Point Cloud Classification with Image-Depth Pre-Training
Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson W.H. Lau, Wanli Ouyang, Wangmeng Zuo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add image-depth contrastive pre-training of the depth encoder (ablation) | CLIP2Point zero-shot · ModelNet10 zero-shot | accuracy (%): 30.51 -> 66.63 | Improved |
| add pre-training (ablation) | CLIP2Point zero-shot · ScanObjectNN OBJ_BG and PB_T50_RS zero-shot | accuracy (%): 23.92 -> 30.46 (OBJ_BG); 18.18 -> 23.32 (PB_T50_RS) | Improved |
| use new depth rendering (minimum depth, dilation) instead of PointCLIP rendering (ablation) | CLIP2Point without pre-training · ScanObjectNN OBJ_ONLY zero-shot | accuracy (%): 15.15 -> 28.40 | Improved |
| add intra-modality contrastive learning to cross-modality pre-training (ablation) | CLIP2Point pre-training · zero-shot classification | accuracy (%): 38.29% -> 11.09% higher | Improved |
| use dual-path with global-view aggregator instead of single path (ablation) | GDPA · ModelNet40 16-shot | accuracy (%): 87.32 -> 89.21 | Improved |
| add gating weights to dual-path fusion (ablation) | GDPA · ModelNet40 16-shot | accuracy (%): 89.21 -> 89.79 | Improved |
| replace inter-view adapter with global-view aggregator (single path) (ablation) | GDPA · ModelNet40 16-shot | accuracy (%): 86.06 -> 87.32 | Improved |
| add dual-path to inter-view adapter (ablation) | GDPA · ModelNet40 16-shot | accuracy (%): 86.06 -> 86.18 | No clear effect |
| initialize both dual paths with CLIP encoder instead of CLIP2Point depth encoder (ablation) | GDPA · ModelNet40 16-shot | accuracy (%): 89.79 -> 87.20 | Hurt |
| use minimum depth value instead of weighted summation (ablation) | depth rendering · ModelNet40 zero-shot (R=2) | accuracy (%): 17.63 -> 29.71 | Improved |
| increase dilation rate from 2 to 4 (ablation) | depth rendering with minimum depth · ModelNet40 zero-shot | accuracy (%): 29.71 -> 28.36 | Hurt |
| replace 2D depth encoder with 3D Point Transformer encoder in pre-training (ablation) | CLIP2Point · ModelNet40 zero-shot | accuracy (%): 49.38 -> 20.83% | Hurt |
Later papers that compared with this method
2 later accepted papers report Clip2point as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- DAP: Doppler-aware Point Network for Heterogeneous mmWave Action Recognition (ECCV 2026 · better on 2 of 2 shared results)
- PointCLIP V2: Prompting CLIP and GPT for Powerful 3D Open-world Learning (ICCV 2023 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.90)
- Duoduo CLIP: Efficient 3D Understanding with Multi-View Images (ICLR 2025 · similarity 0.89)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.89)
- Frozen CLIP Transformer Is an Efficient Point Cloud Encoder (AAAI 2024 · similarity 0.89)
- Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding (ICML 2025 · similarity 0.88)
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (CVPR 2025 · similarity 0.87)
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.
