A Deep Multimodal Model for Predicting Affective Responses Evoked by Movies Based on Shot Segmentation

Predicting the emotions evoked in a viewer watching movies is an important research element in affective video content analysis over a wide range of applications. Generally, the emotion of the audience is evoked by the combined effect of the audio-visual messages of the movies. Current research has mainly used rough middle- and high-level audio and visual features to predict experienced emotions, but combining semantic information to refine features to improve emotion prediction results is still not well studied. Therefore, on the premise of considering the time structure and semantic units of a movie, this paper proposes a shot-based audio-visual feature representation method and a long short-term memory (LSTM) model incorporating a temporal attention mechanism for experienced emotion prediction. First, the shot-based audio-visual feature representation defines a method for extracting and combining audio and visual features of each shot clip, and the advanced pretraining models in the related audio-visual tasks are used to extract the audio and visual features with different semantic levels. Then, four components are included in the prediction model: a nonlinear multimodal feature fusion layer, a temporal feature capture layer, a temporal attention layer, and a sentiment prediction layer. This paper focuses on experienced emotion prediction and evaluates the proposed method on the extended COGNIMUSE dataset. The method performs significantly better than the state-of-the-art while significantly reducing the number of calculations, with increases in the Pearson correlation coefficient (PCC) from 0.46 to 0.62 for arousal and from 0.18 to 0.34 for valence in experienced emotion.