Aspects of the present invention provide an approach for reinforcement learning for unsupervised video summarization with piecewise linear interpolation. A candidate importance score is anticipated by training a video summarization network using a set of video frames in a video. The candidate importance score is interpolated with respect to other candidate importance scores in the video in order to generate a selection probably for each frame of the video. A video summarization that includes a set of selected frames is generated frame selection probability.