INDIVIDUAL HOME-VIDEO COLLECTING USING A CO-CLUSTERING METHOD

摘要

This paper presents a framework to extract subset which only contains one person's video segments from a superfluous home video set. This is a semantic level multimedia application. We proposed a co-clustering method based on facial and body feature, and defined a new type of measurement to combine the two features more reasonable. 2D-PCA detector is used to extract facial feature, and K-Means algorithm is used for clustering. Body features, based on Histograms of Oriented Gradients (HOG) for human detection, combines with Bayes Decision Theory to amend above clustering results. We tested the system in three data sets, including more than 1100 minutes of video. Experimental results show that our approach is feasible and demonstrated good performance in accuracy.