Rule-Based Scene Boundary Detection for Semantic Video Segmentation

In this paper, we present a novel method for semantic video segmentation by using both low-level features and high-level rules on videos and managing it in a hierarchical structure of key-frame, shot and scene. Features in color domain is calculated and utilized for detecting the key-frames and estimating the similarity between shots. By applying the predefined high-level rules, similar shots are merged and the scene boundaries are determined. Finally, a likelihood function is designed for improving the accuracy of scene boundary results. Experimental results from several Hollywood movies have demonstrated and show a better performance of both precision and recall has been achieved comparing with other existing works.