The Paper
Understanding how people actually respond to video content requires more than a self-reported survey score. Organisations need a way to measure attention and emotion directly from behaviour; grounded in real-world data, validated against human judgement, and built to hold up outside the lab.
This technical white paper explains how the Realeyes Attention & Emotion works, from data collection and human annotation through to model architecture, validation results and privacy-preserving design.
What you'll learn
- How attention and emotion are defined as measurable behavioural states, not direct readouts of internal experience
- The scale and diversity of the training dataset: 336 billion viewing data frames from 18 million webcam observations across 90 countries
- How ground truth is established through multi-annotator human agreement before any model is trained
- The CNN, LSTM and TCN architecture behind attention and emotion classification
- How frame-level classifications are aggregated into stable measures like Attention Volume and Attention Quality
- Benchmark validation results against independent human annotators
- How threshold selection and ROC analysis balance sensitivity against false positives
- The privacy-preserving architecture that discards raw imagery and outputs only behavioural classifications
Why this matters
Many attention and emotion models are trained and tested under narrow, ideal conditions, then struggle to generalise once deployed against real devices, lighting and audiences. This whitepaper sets out how a production-grade measurement pipeline is built and validated to perform reliably outside the lab, and is transparent about what the underlying classifiers can and can’t claim to know about a person’s inner state.
Whether you’re evaluating audience research tools, ad effectiveness measurement or engagement analytics, this guide provides the technical detail needed to assess the science behind the numbers.