Perception, Evidence & Efficiency for AI in the Physical World

This is an accompanying post for my ECCV 2026 CONTEXTUS workshop talk.

Workshop: CONTEXTUS @ ECCV 2026
Slides: Download slides

Abstract

To enable AI assistive systems that can operate in physical, real-world environments, we need to go beyond short clips, improve transparency, and enable efficient operation over hours of continuous video. This talk presents three foundational capabilities required for embodied AI systems that work in dynamic, high-stakes settings: Predictive Motion Understanding via UniEgoMotion [1], which anticipates human action from egocentric video to enable perfectly-timed assistance; Evidence-Backed Reasoning via E-VQA [2] , which grounds AI claims in dense spatio-temporal grounding so systems show their work rather than just confidence; and Streaming Efficiency via StateKV [3], which enables real-time inference over long videos.

References

UniEgoMotion: A Unified Model for Egocentric Motion Reconstruction, Forecasting, and Generation

Chaitanya Patel, Hiroki Nakamura, Yuta Kyuragi, Kazuki Kozuka, Juan Carlos Niebles, and Ehsan Adeli

In IEEE/CVF International Conference on Computer Vision (ICCV). Honolulu, Hawaii. Oct 2025

DOI arXiv Bib Video Code Website Data

@inproceedings{Patel_ICCV_2025,
  author = {Patel, Chaitanya and Nakamura, Hiroki and Kyuragi, Yuta and Kozuka, Kazuki and Niebles, Juan Carlos and Adeli, Ehsan},
  title = {{UniEgoMotion}: A Unified Model for Egocentric Motion Reconstruction, Forecasting, and Generation},
  booktitle = {IEEE/CVF International Conference on Computer Vision (ICCV)},
  address = {Honolulu, Hawaii},
  month = oct,
  year = {2025},
  doi = {10.1109/ICCV51701.2025.00961},
}

Evidence-Backed Video Question Answering

Shijie Wang, Honglu Zhou, Ziyang Wang, Ran Xu, Caiming Xiong, Silvio Savarese, Chen Sun, and Juan Carlos Niebles

In European Conference on Computer Vision (ECCV). Malmo, Sweden. Sep 2026

arXiv Bib Code Data

@inproceedings{WangECCV2026,
  title = {Evidence-Backed Video Question Answering},
  author = {Wang, Shijie and Zhou, Honglu and Wang, Ziyang and Xu, Ran and Xiong, Caiming and Savarese, Silvio and Sun, Chen and Niebles, Juan Carlos},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2026},
  month = sep,
  address = {Malmo, Sweden},
}

Linear Scaling Video VLMs for Long Video Understanding

Cristobal Eyzaguirre, Jiajun Wu, and Juan Carlos Niebles

In European Conference on Computer Vision (ECCV). Malmo, Sweden. Sep 2026

arXiv Bib Blog Code

@inproceedings{EyzaguirreStateKV2026,
  title = {Linear Scaling Video VLMs for Long Video Understanding},
  author = {Eyzaguirre, Cristobal and Wu, Jiajun and Niebles, Juan Carlos},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2026},
  month = sep,
  address = {Malmo, Sweden},
}

Abstract

References

Enjoy Reading This Article?