Yuxiong Wang
University of Illinois Urbana-Champaign, IL·Computer Science
Most AI systems that 'watch' video can handle only short clips. Many real uses involve video that runs from several minutes to hours, such as footage from wearable cameras, phones or fixed security cameras, whether live or recorded.
- Fields
Computer Science
Engineering
- Topics
computer vision
video understanding
vision-language models
- Grants
- 1 known grant, through Aug 2028
What they work on
- Develop a multi-resolution video encoding strategy that adjusts token granularity to balance visual detail against efficiency in transformer-based vision-language models.
- Build a selection process aware of content and user intent that keeps a small, informative subset of video tokens and filters out redundancy, so hours-long videos fit within model context limits.
- Create a reasoning module that iteratively explores long video content to support multi-step analysis, temporal reasoning, object grounding and open-ended question answering.
More about the research
Modern vision-language models, AI systems that connect images with text, break video into small units called tokens. A long video can produce millions of tokens. That is far more than these transformer-based models can hold in their working memory (context length), so they struggle to reason over long stretches of time.
Why it matters
Better understanding of long videos could improve interactive AI assistants, autonomous navigation, augmented reality and automatic video summarization. All of these must make sense of minutes to hours of footage from wearable, mobile or fixed cameras, often as it streams live.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Islam Ebeid
Texas Woman's University, TX·Computer Science
Develop an auto-prompting mechanism that uses graph-based reasoning to generate segmentation-specific text descriptions of medical images without human-written prompts.
Rui Yu
University of Louisville, KY·Computer Science
Extract a labeled structural skeleton from single noisy images by fitting geometric primitives such as planes and cylinders to monocular depth estimates, using large vision-language models to assign meaning.
Carl Vondrick
Columbia University, NY·Computer Science·Early-career
Build open-world recognition models for satellite, aerial and ground imagery from vision-language models without language annotations, using cross-modal contrastive learning and reasoning about change over time.
The school
University of Illinois Urbana-Champaign: 41 labs on PI Pages, $907M a year on research, 870 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.