Is this Harmful? Learning to Predict Harmfulness Ratings from Video

Automatically identifying harmful content in video is an important task with a wide range of applications. However, due to the difficulty of collecting high-quality labels as well as demanding computational requirements, the task has not yet had a fully general approach. Typically, only small subsets of the problem are considered, such as identifying violent content. In cases where the general problem is tackled, approximations and simplifications are made to deal with the lack of labels and computational complexity. In this work, we identify and tackle some of the main obstacles. First, we create an open dataset of 3589 video clips from film trailers and annotated by professionals in the field. Second, we perform an analysis of our constructed dataset, investigating among other things the relation between clip and trailer level annotations. Lastly, we train audiovisual models on our dataset and conduct an in-depth study on our modeling choices. We find that results greatly improve by combining the visual and audio modality and that pre-training on large-scale video recognition datasets as well as class balanced sampling further improves performance. Further details of our dataset is available at this webpage: https://vidharm.github.io/.