A MID-LEVEL VISUAL CONCEPT GENERATION FRAMEWORK FOR SPORTS ANALYSIS (WedPmPO1)

Author(s) : Xiaofeng Tong (Chinese Academy of Sciences, China) Rogier Brussee (Telematica Instituut, Netherlands) Lingyu Duan (Institute for Infocomm Research, Singapore) Hanqing Lu (Chinese Academy of Sciences, China) Changsheng Xu (Institute for Infocomm Research, Singapore) Qi Tian (Institute for Infocomm Research, Singapore)

Abstract : A mid-level concept plays a critical role as bridging the gap between low-level feature and high-level semantics in video analysis. Most existing work used to customize special mid-level concepts incorporated with statistics models to detect particular events. Based on existing work, we propose a unified framework of mid-level concept generation for broadcast sports video analysis in this paper. We handle a segment as viewed from three essential and vital aspects: shot size of captured camera, object in the scene and produc-tion way. The three-factor mode clearly illuminates the pri-mary motivation or patterns for a concept generation. With the framework, we can deduce many meaningful mid-level concepts that cover almost all mid-level representations in previous work. Taking these concepts as observations, we can do more comprehensive and interesting video analysis, such as replay classification, detection of goal, shoot, attack, foul, offside, and out of bound, etc.

Menu