Sufficient frame subsets for video understanding

Explaining video classifiers by finding the smallest set of frames that drives their decision.

Video models look at dozens of frames, but which ones actually matter? This project develops an explainable AI method based on sufficient input subsets: it isolates the minimal set of frames that is enough for a model to reach its decision. That helps teams debug and audit video models.

The paper is under review at WACV 2027.