Decoding viewer emotions in video ads.

Area of Science:

• Computer Science
• Artificial Intelligence
• Affective Computing

Background:

• Predicting viewer emotions in videos is crucial for applications like content recommendation and advertising.
• A significant challenge has been the scarcity of large-scale datasets with emotional annotations for videos.
• Existing methods are limited by data availability and the complexity of human emotional responses.

Purpose of the Study:

• To develop and validate a deep learning model for predicting viewer emotional responses to video content.
• To address the data limitations in video emotion understanding by creating a large-scale annotated dataset.
• To provide an accurate and accessible tool for analyzing emotions in video advertisements.

Main Methods:

• Utilized a dataset of over 30,000 video advertisements with approximately 2.3 million emotional annotations from 75 viewers per ad.
• Employed a convolutional neural network integrating both video and audio data to predict emotions in 5-second clips.
• Collected annotations across eight emotion categories: anger, contempt, disgust, fear, happiness, sadness, surprise, and neutral, noting their temporal onset.

Main Results:

• The model achieved an average balanced accuracy of 43.6% in predicting salient 5-second emotional clips.
• Demonstrated high performance in detecting specific emotions, with 55.8% accuracy for happiness and 60.2% for sadness.
• Attained a strong average Area Under the Curve (AUC) of 75% when applied to full advertisements to determine emotional undertones.

Conclusions:

• The study successfully overcomes previous data limitations in video emotion understanding.
• The developed deep learning model provides an accurate solution for predicting viewer emotions in videos.
• The trained networks are available for research purposes, fostering further advancements in the field.