Research map / Point clouds / Point cloud classification
CLIP-GS: Unifying Vision-Language Representation with 3D Gaussian Splatting
Siyu Jiao, Haoye Dong, Yuyang Yin, Zequn Jie, Yinlong Qian, Yao Zhao, Humphrey Shi, Yunchao Wei
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 |
|---|---|---|---|
| fine-tune Uni3D on P&C gaussian attributes (ablation) | Uni3D · Objaverse-GS | Top1 accuracy: +13.3 Top1, +16.2 Top3, and +15.8 Top5 | Improved |
| feed all 3DGS attributes into the network instead of P&C (ablation) | fine-tuned Uni3D · Objaverse-GS | Top1 accuracy: 46.9 -> 44.8 | Hurt |
| add GS Tokenizer (ablation) | fine-tuned Uni3D · Objaverse-GS | Top1 accuracy: 33.6->47.9 Top1, 52.3->69.9 Top3, and 60.1->76.8 Top5 | Improved |
| add image voting loss (ablation) | CLIP-GS with GS Tokenizer · Objaverse-GS | Top1 accuracy: 47.9 -> 48.5 | Improved |
| use multiple ordering strategies instead of one (ablation) | CLIP-GS · Objaverse-GS | Top1 accuracy: 48.3 -> 48.5 | No clear effect |
| initialize from point cloud pretrained weights instead of 2D EVA-CLIP weights (ablation) | CLIP-GS · Objaverse-GS | Top1 accuracy: 33.9 -> 48.5 | Improved |
| unfreeze point cloud pretrained transformer weights (ablation) | CLIP-GS · Objaverse-GS | Top1 accuracy: 46.6 -> 48.5 | Improved |
| scale model from Tiny to Large (ablation) | CLIP-GS · Objaverse-GS | Top1 accuracy: 45.9 -> 48.8 | Improved |
| render 3DGS for 20,000 iterations with SH degree 3 instead of 5,000 iterations with SH degree 0 (ablation) | 3DGS reconstruction · reconstruction quality | PSNR: 34.6 -> 37.1 | Improved |
| use 3DGS input instead of point cloud (Uni3D) | CLIP-GS · Objaverse-GS zero-shot | Top1 accuracy: 47.7 -> 48.5 | Improved |
| use 3DGS input instead of point cloud (Uni3D) | CLIP-GS · Text->3D retrieval | R@1: 27.8 -> 36.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PointRWKV — PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning (AAAI 2025) · better on 1 of 1 shared result
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
- UniGS: Unified Language-Image-3D Pretraining with Gaussian Splatting (ICLR 2025 · similarity 0.89)
- Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration (CVPR 2025 · similarity 0.88)
- Tackling View-Dependent Semantics in 3D Language Gaussian Splatting (ICML 2025 · similarity 0.87)
- Duoduo CLIP: Efficient 3D Understanding with Multi-View Images (ICLR 2025 · similarity 0.87)
- CLIPGaussian: Universal and Multimodal Style Transfer Based on Gaussian Splatting (NeurIPS 2025 · similarity 0.87)
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 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.
