DeepSVU: Towards In-depth Security-oriented Video Understanding via Unified Physical-world Regularized MoE Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Ara
Yujie Jin, Wenxin Zhang, Jingjing Wang, Guodong Zhou · 2026
Read DeepSVU: Towards In-depth Security-oriented Video Understanding via Unified Physical-world Regularized MoE Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Ara by Yujie Jin, Wenxin Zhang, Jingjing Wang, Guodong Zhou (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-18019.txt.
Description
Yujie Jin and colleagues push security-oriented video understanding beyond threat spotting. Earlier work mostly detects and locates threats such as shootings and robberies in video, but cannot explain or judge what caused them. Their new chat-paradigm task, DeepSVU, asks systems to identify and locate threats and also attribute and evaluate the causes behind threatening segments.
Two challenges shape the task: modeling coarse-to-fine physical-world information, from human behavior and object interactions to background context, and balancing those factors adaptively. Their Unified Physical-world Regularized MoE approach pairs two parts: a Unified Physical-world Enhanced MoE block that models that information, and a Physical-world Trade-off Regularizer that balances it. Tests on two DeepSVU instruction datasets show the approach beating advanced video language models and non-vision-language approaches, confirming the value of coarse-to-fine physical-world information.
Catalog record
| CREATOR | Yujie Jin, Wenxin Zhang, Jingjing Wang, Guodong Zhou |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-18019