Research map / Point clouds / Point cloud classification
PointAlign: Feature-Level Alignment Regularization for 3D Vision-Language Models
Yuanhao Su, Shaofeng Zhang, Xiaosong Jia, Qi Fan
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 feature-level alignment loss at LLM layer 16 (lambda 0.1 vs 0) (ablation) | MiniGPT-3D Stage 2 · ModelNet40+Objaverse classification | average accuracy (%): 64.72 -> 66.08 | Improved |
| add PointAlign over MiniGPT-3D on Objaverse classification | MiniGPT-3D · Objaverse classification (instruction prompt) | accuracy (%): 65.00 -> 72.50 | Improved |
| add PointAlign over MiniGPT-3D on ModelNet40 | MiniGPT-3D · ModelNet40 classification (average) | accuracy (%): 61.24 -> 61.17 | No clear effect |
| add PointAlign to MiniGPT-3D for captioning | MiniGPT-3D · Objaverse captioning | Qwen2-72B-Instruct score: 48.17 -> 53.05 | Improved |
| add PointAlign to MiniGPT-3D for captioning | MiniGPT-3D · Objaverse captioning | Sentence-BERT / SimCSE: S-BERT 49.54 -> 49.94; SimCSE 51.39 -> 51.32 | No clear effect |
| use cosine loss instead of L1 or L2 for alignment (ablation) | PointAlign · classification average | accuracy (%): L1 65.72 / L2 65.57 -> cosine 66.08 | Improved |
| align at layer 16 rather than layers 8-32 or multi-layer (ablation) | PointAlign · classification average | accuracy (%): layer 12: 65.14, layer 28: 64.81 -> layer 16: 66.08 | Improved |
| increase alignment weight lambda from 0.1 to 0.3-0.9 (ablation) | PointAlign · classification average | accuracy (%): 66.08 -> 64.39 to 65.08 | Hurt |
| use 3 linear layers in the alignment projector instead of 1, 2 or 4 (ablation) | PointAlign · classification average | accuracy (%): 1 layer 64.45; 2 layers 65.48; 4 layers 65.24 -> 3 layers 66.08 | Improved |
| use Q-Former output as alignment target instead of projector features (ablation) | PointAlign · classification average | accuracy (%): 65.09 (Projector-mid), 64.91 (Projector) -> 66.08 | Improved |
| add alignment regularization when training data fraction grows from 50% to 100% (ablation) | MiniGPT-3D captioning · Objaverse captioning (Qwen2 score) | score: baseline 52.72 -> 48.17 (100%); aligned 53.05 at 100% | Improved |
| add alignment with only 10% of training data (ablation) | MiniGPT-3D captioning · Objaverse captioning (10% data) | Qwen2 score: 41.77 -> 41.95 | No clear effect |
Closest papers by meaning
- More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding (AAAI 2025 · similarity 0.90)
- Context-Aware Alignment and Mutual Masking for 3D-Language Pre-Training (CVPR 2023 · similarity 0.90)
- 3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models (NeurIPS 2026 · similarity 0.89)
- LIBA: Language Instructed Multi-granularity Bridge Assistant for 3D Visual Grounding (AAAI 2025 · similarity 0.88)
- G^3-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding (CVPR 2024 · similarity 0.87)
- From Thousands to Billions: 3D Visual Language Grounding via Render-Supervised Distillation from 2D VLMs (ICML 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.
