Methodology
Three models learned from human data - eye-movement recordings, emotion labels and complexity ratings. Foresight applies them to your creative, turns the predictions into KPIs and positions most KPIs against a reference library.
What it is
Foresight answers three questions about a creative, each with its own model:
| Question | Model | Learned from | Feeds |
|---|---|---|---|
| Where are people likely to look? | Attention model | Recordings of where real people looked while freely viewing images | Heatmap, fogmap, gaze plot, Attention, Clarity, Focus, Spread, areas of interest |
| Which emotions is it likely to evoke? | Emotion model | Images people labelled with the emotion they evoked | The eight emotions, Valence, Emotional Engagement |
| How visually complex does it feel? | Complexity model | People's ratings of how complex images look | Mental Demand |
The E-Score then brings the emotion route and the attention route together.
What feeds it
All three models look at the pixels of your creative. The emotion model also uses a short description of the creative's content as context, so it can take into account what is shown and what the visible text says.
If you pick a platform, the attention and complexity models see your creative inside a simplified version of that platform's screen. The emotion model always sees the original upload. A small, fixed contribution, measured on the platform's empty interface, is then added to the emotion results.
How the models learned
Attention
The attention model learned from eye-tracking research: large collections of images for which researchers recorded where many real people looked while viewing them freely. From these recordings it learned which visual features tend to draw the eye: faces, text, contrast, colour, composition and position in the frame.
The model used in Foresight is optimised to run fast enough to analyse every creative, including long videos. For your creative it predicts a map of how likely each area is to be looked at by a typical viewer during free viewing.
Several views are built on that map:
- Hot spots: Foresight finds the distinct areas where attention clusters. One dominant hot spot guides the eye; several far-apart ones mean attention is divided. They feed Spread and the gaze plot.
- Gaze plot: the hot spots joined in a suggested order, derived from many simulated viewers who start, as people typically do, near the middle of the image. The order is indicative only.
- Attention build-up: a replay of how attention is predicted to develop, from the centre towards the most attention-grabbing areas, until it settles into the final heatmap.
- Areas of interest: the share of predicted attention inside areas you draw, with an estimated viewing time that moves with the share.
Emotion
The emotion model learned from a large collection of images, each labelled by people with the emotion it evoked in them.
For your creative it returns a strength for each of eight emotions. Each value is scored independently, so the eight values don't add up to 100. It predicts the feeling the content evokes in a viewer; it does not read the facial expressions of people shown in the creative or of anyone watching it.
Complexity
The complexity model learned from studies in which people rated how visually complex many different images felt. It predicts how complex a typical viewer would judge your creative. It looks only at what is visible: it does not read the story, the script or the meaning of the text.
How emotions are described
Psychology describes emotions in two complementary ways, and Foresight uses both:
- Discrete view: named emotions. Foresight predicts eight emotions: four pleasant (amusement, awe, contentment, excitement) and four unpleasant (anger, disgust, fear, sadness).
- Dimensional view: any feeling placed on two axes, valence (how pleasant or unpleasant) and arousal (how calm or activating).
Each of the eight emotions has a fixed position on both axes, set by Foresight based on Russell's circumplex model of affect.
Foresight predicts the eight emotions, then derives the dimensional view from them. Emotional Engagement reflects how activating the predicted emotions are; Valence reflects how pleasant they are. Both are derived the same way for every creative. Valence summarises the overall tone. Emotional Engagement summarises intensity regardless of tone, so a strong negative reaction counts like a strong positive one of similar energy. See Valence and intensity.
From predictions to scores
Most Foresight KPIs are rankings. Your creative's raw measurement is placed within a library of reference creatives, and the score says what share of that library it beats: 50 is the middle, and higher scores beat more of it. Focus, Spread and Emotional Engagement work this way, and Clarity is built from Focus and Spread. Mental Demand is also shown as a ranking, but it is best in the middle: content that is far too simple or far too complex is flagged. Valence is the exception: it is not a ranking but the emotional tone of the content itself, from −100 to +100.
The reference library is analysed in the same way as your creative, so both sides are predictions from the same models. The comparison value shown next to a score is the average of the group you selected (for example a sector). See How scores are positioned.
Research foundations
Foresight's models and scores build on published research:
- Emotion training data. Yang, J., Huang, Q., Ding, T., Lischinski, D., Cohen-Or, D. & Huang, H. (2023). EmoSet: A Large-scale Visual Emotion Dataset with Rich Attributes. arXiv:2307.07961.
- Emotion axes. Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology 39(6):1161–1178. DOI 10.1037/h0077714.
- Central viewing bias. Tatler, B. W. (2007). The central fixation bias in scene viewing: Selecting an optimal viewing position independently of motor biases and image feature distributions. Journal of Vision 7(14):4. DOI 10.1167/7.14.4.
- Average fixation statistics. Schwetlick, L., Kümmerer, M., Bethge, M. & Engbert, R. (2024). Potsdam data set of eye movement on natural scenes (DAEMONS). Frontiers in Psychology 15:1389609. DOI 10.3389/fpsyg.2024.1389609.
- Simulated viewing paths. Boccignone, G. & Ferraro, M. (2004). Modelling gaze shift as a constrained random walk. Physica A 331(1–2):207–218. DOI 10.1016/j.physa.2003.09.011. Boccignone, G. & Ferraro, M. (2013). Feed and fly control of visual scanpaths for foveation image processing. Annals of Telecommunications 68(3–4):201–217. PDF.
Image vs video
The same models analyse images and videos. Videos are analysed as a sequence of frames:
- Attention and complexity are predicted on frames sampled throughout the video, and the moments in between are filled in. No motion or sound cues are used.
- Emotion is predicted frame by frame. The video's overall emotion profile is the average over time.