Research map / Vision-language and multimodal / Multimodal LLMs
Stream3D-VLM: Online 3D Spatial Understanding with Incremental Geometry Priors
Hanxun Yu, Xuan Qu, Lei Ke, Boqiang Zhang, YUXIN WANG, Jianke Zhu, Dong Yu
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 geometry priors (camera + geometry tokens) via cross-attention fusion (ablation) | Stream3D-VLM-7B visual baseline · Stream3D-Bench | NA: 46.0 -> 59.8 | Improved |
| add geometry priors to visual baseline (ablation) | Stream3D-VLM · VSI-Bench NA | accuracy: 42.9 -> 65.7 | Improved |
| use cross-attention fusion instead of addition (ablation) | Stream3D-VLM · Stream3D-Bench NA | accuracy: 57.6 -> 59.8 | Improved |
| use cross-attention fusion instead of addition (MCA) (ablation) | Stream3D-VLM · Stream3D-Bench MCA | accuracy: 65.8 -> 65.4 | No clear effect |
| use cross-attention instead of concat + MLP (ablation) | Stream3D-VLM · Stream3D-Bench NA | accuracy: 53.5 -> 59.8 | Improved |
| use GAVC instead of average pooling at 50% retention (ablation) | Stream3D-VLM · Stream3D-Bench NA | accuracy: 47.8 -> 59.8 | Improved |
| use GAVC instead of VisionZip at 50% retention (ablation) | Stream3D-VLM · ScanQA | CIDEr: 87.2 -> 104.5 | Improved |
| replace standard CE with focal loss for streaming tokens (ablation) | Stream3D-VLM · Stream3D-Bench | Answer-Timing Accuracy: 86.7% -> 77.4% | Hurt |
| raise streaming loss weight lambda from 2.0 to 3.0 (ablation) | Stream3D-VLM · Stream3D-Bench | NA: 59.8 -> 57.0 NA; timing 86.7% -> 89.2% | Mixed |
| lower streaming loss weight lambda from 2.0 to 1.0 (ablation) | Stream3D-VLM · Stream3D-Bench | Answer-Timing Accuracy: 86.7% -> 80.3% | Hurt |
| online streaming model vs best offline spatial reasoning model on Relative Direction | Stream3D-VLM-8B · VSI-Bench Rel. Dir. | accuracy: 80.5 -> 73.3 | Hurt |
| use Stream3D-VLM-8B instead of VLM-3R-8B | Stream3D-VLM · VSI-Bench | average accuracy: 60.9 -> 65.9 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- VLM-3R-8B — VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026) · better on 1 of 2 shared results
- Spatial-MLLM-4B — Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence (NeurIPS 2025) · better on 1 of 1 shared result
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
- Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · similarity 0.90)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.90)
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.90)
- EA3D: Online Open-World 3D Object Extraction from Streaming Videos (NeurIPS 2025 · similarity 0.89)
- Video-3D LLM: Learning Position-Aware Video Representation for 3D Scene Understanding (CVPR 2025 · similarity 0.89)
- FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometry (NeurIPS 2026 · 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.
