Video scene segmentation by improved visual shot coherence

Abstract

Nowadays, there a increasing interest in video scene segmentation due huge amount of videos available through services like YouTube. Although there are some techniques which obtain relatively good precision and recall values when segmenting the video in scenes, they are somewhat limited because the high computational cost. A well know technique to accomplish video scene segmentation is the shot coherence model, which presents lower precision and recall than state of art methods, like machine learning and multimodality, but stands out for being simple. The improvement of the techniques based on shot coherence models could be beneficial to these state of the art segmentation methods. That way, this paper presents a new technique for scene segmentation using shot coherence and optical flow features. The technique is presented and evaluated through a series of precision, recall and F1 values, obtaining results close or even better of those obtained by related works.