Indistinct Segmentation of Scene in Video Using Instance Learning

Abstract

In this paper, we indicate that scene boundaries sometimes are indistinct so that computer cannot output explicit results. To solve such problem, we propose that video scenes should be divided into two kinds: cut segmentation and indistinct segmentation and a tolerable value should be given before comparing performance of different algorithms according to different applications. In order to divide two kinds of scenes, we introduce a Fuzzy function to assess ambiguous degree of scene boundary. This paper also present a novel two-pass approach of scene segmentation which is based on constructing temporal graph and Instance learning algorithm. In pass-one, the method first constructs shot temporal directed graph and splits graph into sub-graphs, some sub-graphs are identified as training examples (TEs) by analyzing their density and the nearest neighbor classifier is generated to label shot as-1, 0 or 1. In pass-two, a sequence segmentation algorithm is applied to detect scene boundaries on label sequence. Experiments are presented with promising results on several movies and TV plays.