Research map / Video / Video-language understanding
Towards Spatial Supersensing in Video
Shusheng Yang, Jihan Yang, Pinzhi Huang, Ellis L Brown II, Zihao Yang, Yue Yu, Shengbang Tong, Zihan Zheng, Yifan Xu, Muhan Wang, Rob Fergus, Yann LeCun and 2 more
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 |
|---|---|---|---|
| train on full VSI-590K mixture instead of the LLaVA-Video-178K subset baseline (ablation) | improved Cambrian-1 MLLM · VSI-Bench | average accuracy: 28.5 -> 63.2 | Improved |
| add annotated real-video source ScanNet to the baseline data (ablation) | improved Cambrian-1 MLLM · VSI-Bench | average accuracy: 28.5 -> 56.3 | Improved |
| add pseudo-annotated YouTube room-tour images to the baseline data (ablation) | improved Cambrian-1 MLLM · VSI-Bench | average accuracy: 28.5 -> 32.5 | Improved |
| add configuration question group to VSI-590K training (ablation) | Cambrian-S · VSI-Bench | average accuracy: 51.9 -> 63.2 | Improved |
| add measurement question group to VSI-590K training (ablation) | Cambrian-S · VSI-Bench | average accuracy: 49.7 -> 63.2 | Improved |
| add spatiotemporal question group to VSI-590K training (ablation) | Cambrian-S · VSI-Bench | average accuracy: 58.1 -> 63.2 | Improved |
| switch from uniform 128-frame sampling to 1 FPS sampling on VSR | Cambrian-S-7B · VSI-SUPER Recall | accuracy: 26.7 -> 38.3 at 10 min; 23.3 -> 6.0 at 60 min | Mixed |
| scale the LLM from 0.5B to 7B in Cambrian-S | Cambrian-S · VSI-Bench | accuracy: 50.6 -> 67.5 | Improved |
| add surprise-driven memory (latent-frame-prediction error) to Cambrian-S for VSR (ablation) | Cambrian-S · VSI-SUPER Recall | accuracy | Improved |
| use LFP prediction error as surprise instead of adjacent-frame feature difference (VSR) (ablation) | Cambrian-S memory system · VSI-SUPER Recall | accuracy | Improved |
| use LFP prediction error as surprise instead of adjacent-frame feature difference (VSC) (ablation) | Cambrian-S event segmentation · VSI-SUPER Count | MRA | Improved |
| add surprise-driven event segmentation to Cambrian-S for VSC | Cambrian-S · VSI-SUPER Count | MRA: 38% MRA at 10 min, about 28% at 120 min | Improved |
Closest papers by meaning
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.89)
- DELTAVID: Enhancing Fine-Grained Spatiotemporal Perception with Cross-Video Differences (NeurIPS 2026 · similarity 0.87)
- SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.87)
- Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces (CVPR 2025 · similarity 0.87)
- VideoAnchor: Reinforcing Subspace-Structured Visual Cues for Coherent Visual-Spatial Reasoning (ICLR 2026 · similarity 0.87)
- Exploring Spatial Intelligence from a Generative Perspective (CVPR 2026 · similarity 0.87)
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.
