Researchers have introduced Sekai2, a new large-scale video dataset designed for training world models in AI. This dataset comprises over 128,000 clips totaling 2,826 hours, collected from 10,428 videos across 113 countries. Sekai2 uniquely provides synchronized camera trajectories and detailed hierarchical annotations, disentangling subject motion, environmental dynamics, and scene content. A key feature is the inclusion of 982 panoramic sequences with revisits to the same locations across different times and viewpoints, which is crucial for developing AI that can learn persistent scene representations and long-term spatial memory. AI
IMPACT Enables training of AI world models capable of long-horizon generation and camera control, advancing interactive AI.
RANK_REASON The item describes a new dataset released with a research paper for AI training. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- Connected Papers
- DagsHub
- Gotit.pub
- Hugging Face
- Litmaps
- ScienceCast
- scite Smart Citations
- Sekai2
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
