Automatic Measurement of Visual Attention to Video Content using Deep Learning
2019 16th International Conference on Machine Vision Applications (MVA)(2019)
摘要
Advances in automated face analysis have made possible webcam-based assessment of viewer emotion during presentation of commercials and other video content. A key assumption of this technology is that viewer emotion is in response to the media. Is that assumption warranted? Because viewer attention is seldom assessed, emotional responses could result from other sources, such as talking to a friend, enjoying a meal, or attending to a pet. We developed a CNN-LSTM approach that detects attention and nonattention to commercials using webcam and mobile devices in settings of viewer's choice. Because cultural variation in viewer response is likely, we included participants from both Western and Eastern countries. Participants were 28,911 adults (ages 18 to 69 years) in Europe, USA, Russia, and China. A total of 15,543 sessions (ca. 6.5 million video frames) was analyzed. Accuracy was quantified using a variety of metrics. Our approach outperformed baseline and achieved moderate to high accuracy that approached that of human annotators.
更多查看译文
关键词
automatic measurement,visual attention,video content,deep learning,automated face analysis,possible webcam-based assessment,viewer emotion,commercials,viewer attention,emotional responses,CNN-LSTM approach,viewer response,video frames
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络