Research map / 3D vision / 3D generation
Scenes as Tokens: Multi-Scale Normal Distributions Transform Tokenizer for General 3D Vision-Language Understanding
Yutao Tang, Cheng Zhao, Gaurav Mittal, Rohith Kukkala, Rama Chellappa, Cheng Peng, Mei Chen
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 |
|---|---|---|---|
| replace downsampled multi-scale points with multi-scale NDT cells (ablation) | NDTokenizer3D · ScanQA | CiDEr: 94.7 -> 98.6 | Improved |
| replace downsampled multi-scale points with multi-scale NDT cells (ablation) | NDTokenizer3D · Multi3DRefer val | mIoU: 45.3 -> 46.0 | Improved |
| use three scales {2,3,4} instead of a single scale (r=3) (ablation) | NDTokenizer3D · Multi3DRefer val | mIoU: 40.1 -> 46.0 | Improved |
| use three scales {2,3,4} instead of a single scale (r=4) (ablation) | NDTokenizer3D · SQA3D | EM-R: 56.9 -> 57.1 | Improved |
| add a fourth, finest scale (r={1,2,3,4}) to the three-scale design (ablation) | NDTokenizer3D · ScanQA | CiDEr: 98.6 -> 94.7 | Hurt |
| add a fourth scale (ablation) | NDTokenizer3D · SQA3D | EM-R: 57.1 -> 57.3 | No clear effect |
| increase learnable queries from 100 to 850 (ablation) | NDTokenizer3D · ScanQA | CiDEr: 90.2 -> 98.6 | Improved |
| increase queries from 850 to 1200 or 1600 (ablation) | NDTokenizer3D · ScanQA | CiDEr: 98.6 -> 96.3 (1200), 95.9 (1600) | Hurt |
| use NDTokenizer3D instead of 3D-LLaVA | generalist 3D VLM · Multi3DRefer val | mIoU: 42.7 -> 46.0 | Improved |
| use NDTokenizer3D instead of 3D-LLaVA | generalist 3D VLM · SQA3D | EM: 54.5 -> 54.4 | No clear effect |
| use NDTokenizer3D instead of 3D-LLaVA | generalist 3D VLM · 3D-POPE (Adversarial) | accuracy: 69.88 -> 72.03 | Improved |
Closest papers by meaning
- Merge3D: Efficient 3D Multimodal LLMs via Joint 2D-3D Token Merging (CVPR 2026 · similarity 0.90)
- Lexicon3D: Probing Visual Foundation Models for Complex 3D Scene Understanding (NeurIPS 2024 · similarity 0.90)
- A Unified Framework for 3D Scene Understanding (NeurIPS 2024 · similarity 0.90)
- MLLMs Need 3D-Aware Representation Supervision for Scene Understanding (NeurIPS 2025 · similarity 0.90)
- GPT4Scene: Understand 3D Scenes from Videos with Vision-Language Models (ICLR 2026 · similarity 0.90)
- MLLM-For3D: Adapting Multimodal Large Language Model for 3D Reasoning Segmentation (NeurIPS 2025 · similarity 0.90)
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.
