Researchers have introduced MoCKEFA, a method for selecting keyframes from videos. According to the arXiv abstract, the central claim is that a video can be reduced to just a few keyframes without sacrificing a given model's belief about the video.
The approach is described as the first to extract keyframes conditioned on the model's prediction. This means the selection is tailored to what the model already believes, rather than using a generic or content-only criterion.
The method is designed for video model explanation, and the title also mentions fair attribution. The abstract does not elaborate on how fair attribution is achieved, but the stated goal is to represent videos efficiently for explanation purposes. As a single preprint, the results have not yet been independently verified.