Violent scenes detection with large, brute-forced acoustic and visual feature sets

This paper describes the TUM approaches for violent scenes detection in movies, submitted for the MediaEval 2012 Affect Challenge. Score fusion is used to fuse Support-Vector Machine (SVM) confidence scores assigned to short fixed length windows within each movie shot. SVM predictors for acoustic and visual channels are trained. For the acoustic channel, a large set of acoustic features based on the set from the INTERSPEECH 2012 Speaker Trait Challenge is employed. A comprehensive set of common video low-level descriptors such as optical flow, gradients, and hue and saturation histograms is used for the visual channel.