Research map / Point clouds / Point cloud classification
PointLLM-V2: Empowering Large Language Models to Better Understand Point Clouds
Runsen Xu, Shuai Yang, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, Dahua Lin
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 |
|---|---|---|---|
| use a 3-layer projector (hidden 1024, 2048) instead of 1 layer (ablation) | PointLLM-7B · Generative classification average accuracy | accuracy: 51.05 -> 52.82 | Improved |
| use a 3-layer projector instead of 1 layer (ablation) | PointLLM-13B · Generative classification average accuracy | accuracy: 49.00 -> 53.39 | Improved |
| use a 3-layer projector instead of a 4-layer one (ablation) | PointLLM-13B · Generative classification average accuracy | accuracy: 51.40 -> 53.39 | Improved |
| use a 3-layer projector instead of 4 layers (ablation) | PointLLM-7B · Generative classification average accuracy | accuracy: 52.15 -> 52.82 | Improved |
| add max pooling of point tokens before the projector (ablation) | PointLLM-7B · Generative classification average accuracy | accuracy: 52.82 -> 48.72 | Hurt |
| add max pooling of point tokens before the projector (ablation) | PointLLM-13B · Generative classification average accuracy | accuracy: 53.39 -> 51.10 | Hurt |
| add max pooling of point tokens (ablation) | PointLLM-7B · Training cost | A100 GPU-hours: 126 -> 34 | Improved |
| add max pooling of point tokens (ablation) | PointLLM-13B · Training cost | A100 GPU-hours: 213 -> 56 | Improved |
| add multi-round conversations to fine-tuning data (ablation) | PointLLM-7B · Generative classification average accuracy | accuracy: 40.14 -> 45.79 | Improved |
| add detailed descriptions to fine-tuning data (ablation) | PointLLM-7B · Generative classification average accuracy | accuracy: 45.79 -> 52.82 | Improved |
| increase alignment-stage data volume (ablation) | PointLLM-7B · Generative classification average accuracy | accuracy: plateaus at around 600K samples | Improved |
| prompt for captions of at most 20 words | PointLLM-13B · 3D object captioning (Objaverse) | BLEU-1: 3.83 -> 17.09 | Improved |
Closest papers by meaning
- More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding (AAAI 2025 · similarity 0.92)
- 3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models (NeurIPS 2026 · similarity 0.91)
- GPT4Point: A Unified Framework for Point-Language Understanding and Generation (CVPR 2024 · similarity 0.90)
- PerLA: Perceptive 3D Language Assistant (CVPR 2025 · similarity 0.89)
- LiDAR-LLM: Exploring the Potential of Large Language Models for 3D LiDAR Understanding (AAAI 2025 · similarity 0.88)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 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.
