Research map / Vision-language and multimodal / Multimodal LLMs
SNOW: Spatio-Temporal Scene Understanding with World Knowledge for Open-World Embodied Reasoning
Tin Stribor Sohn, Maximilian Dillitzer, Jason Corso, Johannes Bach, Eric Sax
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 training-free 4DSG with STEP tokens (ablation) | Gemma3-4B-IT · VLM4D (200-question subset) | overall accuracy: 49.25% (A1) to 77.5% (A3) | Improved |
| add 2D temporal tracking to VLM-only baseline (ablation) | Gemma3-4B-IT · VLM4D (200-question subset) | egocentric/exocentric accuracy: +18% ego, +20% exo | Improved |
| add 3D STEP tokens on top of 2D tracking (ablation) | Gemma3-4B-IT · VLM4D (200-question subset) | egocentric/exocentric accuracy: +22% ego, +20% exo | Improved |
| add 4D STEP tokens (ablation) | Gemma3-4B-IT · VLM4D (200-question subset) | false-positive accuracy: 74.0 -> 74.0 | No clear effect |
| add 2D temporal tracking (ablation) | Gemma3-4B-IT · VLM4D (200-question subset) | false-positive accuracy: 74.0 -> 72.0 | Hurt |
| add SNOW 4DSG to Gemma3-4B-IT | SNOW · NuScenes-QA | Status accuracy: +23.5% | Improved |
| add SNOW 4DSG | SNOW · NuScenes-QA | overall accuracy: 58.5 -> 60.1 | Improved |
| use SNOW instead of best prior on Comparison | SNOW · NuScenes-QA | Comparison accuracy: 68.8 -> 61.0 | Hurt |
| use training-free SNOW vs RS-finetuned models | SNOW · RoboSpatial-Home | Context accuracy: 31.1 -> 54.92 | Improved |
| use SNOW vs best prior | SNOW · RoboSpatial-Home | Compatibility accuracy: 81.0 -> 78.10 | Hurt |
| use SNOW vs Gemini-2.5-Pro | SNOW · VLM4D | overall accuracy: 62.0 -> 73.75 | Improved |
| use training-free SNOW vs prior trained open-vocab methods | SNOW · NuScenes LiDAR segmentation (val) | mIoU: 38.1 vs 44.6 (OV3D) | Mixed |
Later papers that compared with this method
1 later accepted paper reports SNOW as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- ROSE: Real-Time Open-World Scene Understanding from Monocular Video via Compact Multimodal 4D Scene Graphs (ECCV 2026 · better on 3 of 5 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- OpenDriveVLA-0.5B — OpenDriveVLA: Towards End-to-end Autonomous Driving with Large Vision Language Action Model (AAAI 2026) · better on 4 of 4 shared results
- LiDAR-LLM — LiDAR-LLM: Exploring the Potential of Large Language Models for 3D LiDAR Understanding (AAAI 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
- ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.90)
- OneCanvas: 3D Scene Understanding via Panoramic Reprojection (NeurIPS 2026 · similarity 0.90)
- OST-Bench: Evaluating the Capabilities of MLLMs in Online Spatio-temporal Scene Understanding (NeurIPS 2025 · similarity 0.90)
- RoboSpatial: Teaching Spatial Understanding to 2D and 3D Vision-Language Models for Robotics (CVPR 2025 · similarity 0.89)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.89)
- PanoWorld: Towards Spatial Supersensing in 360◦ Panorama World (NeurIPS 2026 · 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.
