Inflated 3D ConvNet context analysis for violence detection

Abstract According to the Wall Street Journal, one billion surveillance cameras will be deployed around world by 2021. This amount of information can hardly managed humans. Using a Inflated 3D ConvNet as backbone, this paper introduces novel automatic violence detection approach that outperforms state-of-the-art existing proposals. Most those proposals consider pre-processing step only focus on some regions interest in scene, i.e., actually containing human subject. In regard, also reports results an extensive analysis whether and how context affect or not adopted classifier performance. The experiments show context-free footage yields substantial deterioration performance (2% 5%) publicly available datasets. However, they demonstrate stabilizes settings, no matter level restriction applied. Finally, cross-dataset experiment investigates generalizability obtained single-collection (same dataset used for training testing) cross-collection settings (different datasets testing).