A Self-Adjusting Shot-Clustering Technique Without Experiential Parameters
一种不需经验参数的视频镜头自校正聚类方法

Shot clustering is an important issue in the field of video content analysis.The basic task of shot clustering is to classify shots based on their low level physical features.This paper describes a novel shot clustering technique.Beginning with an initial classification of the shot set,our algorithm proceeds with merging and splitting iteration alternatively to reduce the errors in the initial results.The initial classification is based on the closeness of shots.The following clustering process is controlled by merging and splitting rules,which are based on the concepts of centroid,radius,intra cluster hole,and inter cluster space.The basic idea is that:(1)the intra cluster hole of a cluster should be less than the intercluster space; and (2)the centroid distance of two clusters should be larger than the sum of their radiuses.Special consideration is put on the design of the iteration mode to suppress the possible errors.The main advantage of this algorithm is that it does not need any experiential parameters or thresholds,neither does it need any manual interaction,which is a basic requirement for automatic clustering of shots with no domain knowledge.Experimental results are presented and analyzed.