Duygusal patlama sezimi ve tanıması

Affect bursts play a critical role in recognizing underlying emotions and hence their detection is of particular interest to researchers. In the literature, there are a number of approaches for detecting affect bursts, particularly laughters, from audio-only inputs, but multimodal approaches are limited while they can provide more effective systems for the affect burst detection. This thesis presents the current state of the art as well as our approaches to detect affect bursts along with their type. This includes a two layer hierarchical classifier, first to detect affect burst events and then to classify these events into affect burst types. In the hierarchical classifier we combine cues from audio and facial landmark points. In experimental evaluations we use the Interactive emotional dyadic motion capture database(IEMOCAP), which contains realistic and natural dyadic conversations with high quality audio and normalized motion capture of facial landmark points. We firstly annotate affect bursts events in the IEMOCAP database and test our proposed approach over it. We observe significant performance improvements with the multimodal approach over audio-only and visual-only unimodal schemes in detecting affect bursts along with their types. We also show comparison among different types of classifiers which serve as baseline.