Learning from Annotated Video: An Initial Study Based on Oyama Karate Tournament Recordings

The most of the up-to-date video data are amateur films consisted of an unbroken sequence of frames recorded from a single camera. In case of karate many even most important world tournaments do not have video broadcasting and the multimedia materials are mostly unprofessional. This type of recording is very difficult to analyze with automatic annotation methods. It is due the fact that a cameraman films the whole fight from stationary position. Only with a help of experienced specialists it is possible to distinguish the karate techniques fighters perform. Due to this fact annotations have to be done mostly manually. We have observed that none of the state of the art papers deal with problem of martial-arts annotated video analysis. It seems that this kind of data might be a very good source of knowledge about a kumite (a fight), a particular fighter and his or her strategies and abilities. In this paper we propose a novel video annotation method that enables both quantitative (numerical) and qualitative (categorical) features calculation. We also present and discuss example results we can obtain from those descriptions with popular data mining methods.