A semi-automatic editing method for surgery videos

Zhiwei Fan, Congliang Chen, Tingting Jiang

Abstract: The editing of a raw surgery video is expensive and time-consuming, for it can take an editor with professional medical knowledge hours. We investigate the possibility of reducing the editing cost and propose a feasible semi-automatic editing method for surgery videos. With our method, the editor just needs to annotate a very small part of the video segments in the raw video. And then a model is trained with the partially labeled segments, which can be used to generate an edited version of the whole video according to the editor's criterion. An active learning strategy is adopted here to reduce the number of video segments that need to be annotated. To verify the function of our method, we build a dataset of two raw surgery videos with their edited versions. It shows that two edited versions of the same raw video can be very different because of different editing criteria. And simulation experiments show that our method is able to generate an edited video meeting the expected editing criteria with limited human annotations.