Video-level violence rating with rank prediction

Given a video as input, our objective is to estimate a rate to describe "how violent it is". Such an estimation can be directly used in many practical applications, such like preventing children from violent videos. However, due to the unique property of the rating task, existing approaches on human action recognition and violent scenes detection can not be directly utilized. In this paper, we propose an approach that are specially developed for violence rating. The approach is featured with: (1) a novel video descriptor called Violent Attribute Activation (VAA) vector, which provides high level description on the properties of visual violence; and (2) a rank-prediction-based rating approach, which enforces the order constrains in the learning phase. The performance of our approach have been confirmed on a novel dataset that are prepared for violence rating.