REINFORCEMENT LEARNING FOR UNSUPERVISED VIDEO SUMMARIZATION WITH PRECEWISE LINEAR INTERPOLATION

摘要

Disclosed is a reinforcement learning technique for unsupervised video summarization using sub-section linear interpolation. A video summarization method performed by a video summarization system according to an embodiment includes: predicting an importance score candidate through training of a video summarization network using a video frame; Interpolating the predicted importance score candidates to obtain a frame selection probability for each frame of a video; And generating a video summary using the selected frame based on the acquired frame selection probability.