Predictive Eye Tracking: Can AI Pass the Test?

In the world of digital marketing and advertising, understanding what consumers look at and what captures their attention is crucial. Predictive eye tracking technology offers the ability to predict where consumers are likely to look in static images and videos. But how accurate is this technology in 2024? 

Predictive Eye Tracking vs. Real Eye Tracking: Can AI Pass the Test?

Recent developments in machine learning and AI have shown that it is possible to train an AI model to predict where consumers will look in an image. But what happens when we throw videos into the mix? Videos are not just a series of consecutive images; they also contain a story and context. Can predictive eye tracking models hold up against real eye tracking data?

AI: Accurate in Focus Points

Predictive eye tracking proves to accurately predict viewer focus points in about 70% of cases. This is particularly impressive in contexts such as TV commercials, where capturing attention is crucial. The model accurately tracks faces and objects within videos, as shown in Figure 1. This level of accuracy is significant and helps marketers better understand where viewers focus their attention.

 

Figure 1: Predictive eye tracking accurately predicts eye movements, focusing on action and faces (vs. regular eye tracking).

Response Speed of Predictive Eye Tracking

A notable challenge is the model's response speed to scene changes. Often, the model immediately focuses on faces when a new scene begins, as shown in Figure 2. In reality, viewers initially pay attention to other objects in a scene before shifting their focus to faces. This difference highlights a shortcoming of the model in mimicking the natural delay and progression of human attention in dynamic environments.

 

Figure 2: Predictive AI focuses too quickly directly on faces.

Nuances of the Scene

Additionally, predictive eye tracking sometimes misses the subtleties of a scene. For example, it may ignore logos after their continuous presence or not accurately follow the main action, such as pouring milk. Instead, the model may focus too much on static elements like hands and blenders, as shown in Figure 3. These nuances are crucial for a full contextual understanding of the scene.

 

Figure 3: Predictive AI focuses on static objects instead of the main action.

Conclusion: Quickly Identifying Predictable Problems

AI, and particularly predictive eye tracking, is a powerful tool for quickly identifying predictable problems. At Unravel Research, we primarily use this technology for our tool AdVisor. This allows us to quickly check whether brand assets are (likely) seen or not. However, it is important that these brand assets are also genuinely valuable for the brand; otherwise, there is still much work to be done.

Brand assets cannot be predicted by AI because their impact depends on how active they currently are in the consumer's brain. Therefore, good Brand Asset research is still needed. 

The nuances and context of a scene can sometimes be lost when using AI, making it crucial to continue using real eye tracking data to get a complete picture of viewing behaviour.

Predictive eye tracking technology can be a valuable addition to traditional methods, especially for quick iterations of video content. By combining the strengths of AI with real neuro-research, marketers can more effectively respond to the attention and engagement of their target audience.