Researchers at the University of Saskatchewan (USask) have developed a tool to make real-world AI faster and more efficient, for applications in medical procedures to self-driving cars.
“It is very important that AI models make decisions faster and more accurately,” said Tzu-Ling Liu, a recent graduate of USask’s Master of Science degree and one of the paper’s authors. “Imagine your self-driving car being able to identify and detect hazards in 0.1 seconds. But our model can decide to stop in a fraction of that time. Which car would you buy?”
This paper was presented at the Institute of Electrical and Electronics Engineers/Computer Vision Foundation (IEEE/CVF) Computer Vision and Pattern Recognition (CVPR) 2026 Conference. The conference is considered one of the most influential publishing venues in the world.
The paper was written by Liu and co-authors Ian Stavness, PhD, chair of USask’s Department of Computer Science, and Mrigank Rochan, PhD, assistant professor in the department.
With today’s models, if an AI recognizes an action through video data, it can have a hard time recognizing that action once it’s trained in one environment and applied to another. For example, an AI trained to recognize a person running on a sunny street will have difficulty recognizing someone running in a dark, rainy park.
The problem of training an AI model in one environment and performing poorly in another is known as “domain shift.”
The research team at USask has developed a system called “Learnable Motion-Focused Tokenization” (LMFT) that acts as a digital filter for training AI models. This removes unimportant background information and allows the model to focus and learn only on the actions that are happening.
“The background hinders the AI’s ability to identify actions, so imagine being able to remove these uninformative background patches,” Liu said.
