Research map / 3D vision / 3D generation
LLaVA³: Representing 3D Scenes Like a Cubist Painter to Boost 3D Scene Understanding of VLMs
Doriand Petit, Steve Bourgeois, Vincent Gay-Bellile, Florian Chabot, Loïc Barthe
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 |
|---|---|---|---|
| object-based scene decomposition (ablation) | Ray-NeRF · MSR3D test | overall correctness: +6.85 | Improved |
| multi-scale hierarchical sampling (ablation) | LLaVA3 · MSR3D test | overall correctness: +2.5 | Improved |
| filtering segmentation (ablation) | LLaVA3 · MSR3D test | overall correctness: +0.55 | Improved |
| radar sweeping ordering (ablation) | LLaVA3 · MSR3D test | overall correctness: +2.09 | Improved |
| adaptive VI-VD instead of even split (ablation) | LLaVA3 · MSR3D test | overall correctness: 43.84 -> 44.89 | Improved |
| 2D object-based decomposition (SAM masks) on LLaVA-OV | LLaVA-OV · ScanQA val | CIDEr: -3.60 | Hurt |
| NeRF reconstruction instead of 2D multi-view | LLaVA-OV · ScanQA val | CIDEr: +5.18 | Improved |
| joint training of decoder instead of frozen AE (ablation) | LLaVA3 · ScanQA subset (4 scenes) | CIDEr: 79.63 -> 97.82 | Improved |
| pre-train auto-encoder before joint training (ablation) | LLaVA3 · ScanQA subset (4 scenes) | CIDEr: 96.77 -> 97.82 | No clear effect |
| add VI-VD double loss (ablation) | LLaVA3 (100% VI) · ScanQA subset (4 scenes) | CIDEr: 72.73 -> 90.82 | Improved |
| add all three filtering steps (ablation) | LLaVA3 · ScanQA subset (4 scenes) | CIDEr: 91.68 -> 97.82 | Improved |
| add radar sweeping (ablation) | LLaVA3 (no multi-scale) · ScanQA subset (4 scenes) | CIDEr: 82.19 -> 87.67 | 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.
- SplatTalk — SplatTalk: 3D VQA with Gaussian Splatting (ICCV 2025) · better on 3 of 3 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
- Language-Image Models with 3D Understanding (ICLR 2025 · similarity 0.92)
- 3D-LLaVA: Towards Generalist 3D LMMs with Omni Superpoint Transformer (CVPR 2025 · similarity 0.90)
- Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction–Reasoning Synergy (ICLR 2026 · similarity 0.90)
- Video-3D LLM: Learning Position-Aware Video Representation for 3D Scene Understanding (CVPR 2025 · similarity 0.90)
- MMScan: A Multi-Modal 3D Scene Dataset with Hierarchical Grounded Language Annotations (NeurIPS 2024 · similarity 0.90)
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.89)
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.
