Soccer Video Event Detection Using 3D Convolutional Networks and Shot Boundary Detection via Deep Feature Distance

In this work, we propose a novel framework combining temporal action localization and play-break (PB) rules for soccer video event detection. Firstly we treat event detection task in action-level, and adopt 3D convolutional networks to perform action localization. Then we employ PB rules to organize actions into events using long view and replay logo detected in the first step. Finally, we determine the semantic classes of events according to principal actions which contain key semantic information of highlights. For long untrimmed videos, we propose a shot boundary detection method using deep feature distance (DFD) to reduce the number of proposals and improve the performance of localization. Experiment results verify the effectiveness of our framework on a new dataset which contains 152 classes of semantic actions and scenes in soccer video.