Comparative Analysis of Low-Level Visual Features for Affective Determination of Video Clips

Many algorithms and works have helped in the understanding and development of affective analysis of films. In spite of the progress made up to now, it is still not very precise how the low-level features of movies shape the resulting affective state of the viewer. In this work we evaluate different visual features and investigate how they impact the emotional evaluation under two paradigms: the dimensional approach (in terms of Pleasure, Arousal and Dominance) and the categorial approach. The analysis is conducted by using Dynamic Bayesian Networks (DBN) in the following topologies: a Hidden Markov Model network and Auto Regressive Hidden Markov Model network. Different combinations of feature vectors were used in order to check the performance of the proposed methods. The ground truth was created from an extensive user experiment, which was also used to create models that could map Pleasure, Arousal and Dominance values into affective categories.