Fixation Prediction through Multimodal Analysis

In this article, we propose to predict human eye fixation through incorporating both audio and visual cues. Traditional visual attention models generally make the utmost of stimuli's visual features, yet they bypass all audio information. In the real world, however, we not only direct our gaze according to visual saliency; but also are attracted by salient audio cues. Psychological experiments show that audio has an influence on visual attention, and subjects tend to be attracted by the sound sources. Therefbre, we propose fusing both audio and visual information to predict eye fixation. In our proposed framework, we first localize the moving-sound-generating objects through multimodal analysis and generate an audio attention map. Then, we calculate the spatial and temporal attention maps using the visual modality. Finally, the audio, spatial, and temporal attention maps are fused to generate the final audiovisual saliency map. The proposed method is applicable to scenes containing moving sound-generating objects. We gather a set of video sequences and collect eye tracking data under an audiovisual test condition. Experiment results show that we can achieve better eye fixation prediction performance when taking both audio and visual cues into consideration, especially in some typical scenes in which object motion and audio are highly correlated.