A study of using simple features for video classification

Abstract

The subject of video classification is an area that has come into attention, especially with the huge amount multimedia content being produced and uploaded due to the convenience to do so with the help of advanced technology. In particular, videos of television shows is where video classification can be applied on as there is a myriad range of shows, such as dramas to variety shows. It would be desirable to be able to classify or categorize these multimedia contents for further applications such as easier search or retrieval.

This project aims study how the use of using simple object-based visual features will affect the accuracy of video classification results. The first phase of the project involves building up a training set of 4 different variety shows and doing face detection on them. The second phase involves feature extraction from episodes of each show using object-based visual features. The last phase is to do testing with a set of 20 videos to see if they fall correctly into their respective shows. The results turned out to be encouraging, with an average of up to 84.60% accuracy.