Affect in Multimedia: Benchmarking Violent Scenes Detection

In this article, we report on the creation of a publicly available, common evaluation framework for Violent Scenes Detection (VSD) in Hollywood and YouTube videos. We propose robust data set, VSD96, with more than 96 hours video various genres, annotations at different levels detail (e.g., shot-level, segment-level), mid-level concepts blood, fire), pre-computed multi-modal descriptors, over 230 system output results as baselines. This is most comprehensive set available to date tailored VSD task was extensively validated during MediaEval benchmarking campaigns. Furthermore, provide an in-depth analysis crucial components algorithms, by reviewing capabilities evolution existing systems overall trends outliers, influence employed features fusion techniques, deep learning approaches). Finally, discuss possibility going beyond state-of-the-art performance via ad-hoc late approach. Experimentation carried out VSD96 data. important lessons learned gained insights. The increasing number publications using underline importance topic. presented published resources are practitioner's guide also strong baseline overcome, which will help researchers coming years analyzing aspects audio-visual affect violence detection movies