Research map / Vision-language and multimodal / Multimodal LLMs
3DSRBench: A Comprehensive 3D Spatial Reasoning Benchmark
Wufei Ma, Haoyu Chen, Guofeng Zhang, Yu-Cheng Chou, Jieneng Chen, Celso de Melo, Alan Yuille
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 DINOv2 as a second vision encoder (ablation) | LLaVA-v1.5-7B · 3DSRBench | mean accuracy: 36.8 -> 37.2 | Improved |
| add MAE as a second vision encoder (ablation) | LLaVA-v1.5-7B · 3DSRBench | mean accuracy: 36.8 -> 33.1 | Hurt |
| add SAM as a second vision encoder (ablation) | LLaVA-v1.5-7B · 3DSRBench | mean accuracy: 36.8 -> 27.9 | Hurt |
| add SAM as a second vision encoder (ablation) | LLaVA-v1.5-7B · 3DSRBench height | height accuracy: 38.5 -> 44.6 | Improved |
| replace MLP with SVA connector for CLIP+DINOv2 (ablation) | LLaVA-v1.5-7B · 3DSRBench | mean accuracy: 37.2 -> 37.8 | Improved |
| replace MLP with SVA connector for CLIP+MAE (ablation) | LLaVA-v1.5-7B · 3DSRBench | mean accuracy: 33.1 -> 34.1 | Improved |
| scale language model size from 0.5B to 72B (ablation) | Qwen2.5 and InternVL2.5 · 3DSRBench-real | accuracy | Improved |
| use proprietary LMMs instead of open-sourced ones | proprietary LMMs · 3DSRBench-real | overall accuracy: 44.2 vs 50.9 | No clear effect |
| add depth input to SpatialRGPT (ablation) | SpatialRGPT · 3DSRBench-real | overall accuracy: 32.7 -> 48.4 | Improved |
| add explicit thinking to Gemini-2.0-Flash (ablation) | Gemini-2.0-Flash · 3DSRBench-real | overall accuracy: 49.8 -> 51.1 | Improved |
| add 3D bounding-box prompting to Gemini-2.0-Flash (ablation) | Gemini-2.0-Flash · 3DSRBench-real | overall accuracy: 49.8 -> 47.5 | Hurt |
| change camera viewpoint from common to uncommon | GPT-4o · 3DSRBench-synthetic | overall accuracy: 51.2 -> 44.3 | Hurt |
Closest papers by meaning
- SpatialReasoner: Towards Explicit and Generalizable 3D Spatial Reasoning (NeurIPS 2025 · similarity 0.93)
- Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models (CVPR 2025 · similarity 0.93)
- Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.92)
- SpatialLLM: A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models (CVPR 2025 · similarity 0.92)
- SpaceVista: All-Scale Visual Spatial Reasoning from mm to km (ICML 2026 · similarity 0.91)
- Spatial Reasoning with Vision-Language Models in Ego-Centric Multi-View Scenes (ICLR 2026 · similarity 0.91)
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.
