RGVCD: A New Real-time Game Video Clip Detection System

Interesting video clip detection is an important task in computer vision, which can be used as data preprocessing for a wide range of applications such as video retrieval and content-based video clustering. In this paper, we propose a new real-time game video clip detection system for interesting moments termed as RGVCD. During training period, we generate a set of game video clips to train a spatial-temporal network. Given a test video, RGVCD first generates potentially interesting video clips. For each clip, we extract features from images and optical flow using deep convolutional network. Finally, RGVCD outputs interesting degree score based on the spatial-temporal features. Additionally, we collect a set of games video clips and provide the corresponding binary label to denote whether the video clip is of our interest. With this dataset, our system demonstrates its efficiency in real-time interesting video clip detection with high accuracy.