ICME Grand Challenge on Short Video Understanding

In recent years, short video applications have entered our lives and lots of short videos are watched by users every day. ICME and Byte-Dance corporation sponsor the grand challenge for better understanding video content and recommending what users like. The challenge provides an incremental multi-modal video dataset including face features, title features, video content features, audio content features and users' behaviors features consisting of thousands of different users and millions of different videos. This paper reports the implementation details in the grand challenge. We have made different attempts and explorations in feature engineering and model structure, finally we got good results and achievements. The details of feature design, model experiments and ensemble learning are described. In this paper, we provides the implementation details of the whole challenge.