Eye tracking shows how quickly faces draw attention in paintings

Faces captured viewers’ attention quickly in a study of how people look at paintings. Among works containing faces, the median first fixation on a face arrived after 0.39 seconds. Those faces occupied 5.1% of the relevant image area but received 20.1% of fixations.[1]

The AVAtt preprint reports 4,505 viewing paths and 121,426 fixations from 152 participants looking at 283 paintings across 16 art movements. For AI researchers, it offers a way to test whether models largely developed on photographs can predict attention in images deliberately composed by artists.[1]

The pull of a painted face

The face measurements were compared with a baseline designed to account for viewing biases. The baseline put the median first face fixation at 3.92 seconds and the share of viewers reaching a face within one second at 17%, compared with an observed 57%. These are measurements of looking, not a score for how much a painting was liked.[1]

The researchers also found that people's scanning paths varied by style and subject. In some paintings, the gaze followed prominent edges, colour contrasts or clusters of objects. In others, a person or animal anchored the search. A saliency model that can predict where viewers look in a street photograph may not correctly anticipate how a painter arranges competing points of emphasis.[1]

Can a neural network learn to look at art?

The team compared computational models with measured human attention. Some models produced respectable estimates of likely fixation density, but they did not match the agreement achievable between human viewers under the reported metrics. One benchmark comparison placed the strongest model at a correlation of around 0.695 while the human upper-reference value reached approximately 0.807. Those figures depend on the metric, the dataset and the precise evaluation protocol; they should not be read as a general intelligence contest between people and machines.[1]

Why is painting difficult? First, artistic images can contain exaggerated forms or invented lighting that violates the assumptions a model learns from camera images. Second, viewers may use cultural context and learned expectations. Third, the signal that matters is dynamic: predicting a hot region on a heat map is not the same as predicting a precise sequence of eye movements through time.[1]

The dataset's value is that it makes those differences measurable. A model can be tested against observed locations and scanpaths, rather than praised because its heat map happens to look plausible. Researchers can also ask whether a technique performs differently on abstract paintings, portraits or landscapes, and whether the difference persists when the paintings themselves change.[1]

From a heat map to a sequence of looks

The participant group has limitations. Most were novices rather than trained art historians, and recruitment was geographically concentrated. They looked at reproductions under experimental conditions, not original paintings in museums where scale, physical texture, lighting and time spent in front of the work may change behaviour. The paper also extends earlier research, so it is an expanded dataset and benchmark, not the first observation that human eyes favour faces.[1]

A useful model must distinguish two tasks: predicting which regions attract attention overall and predicting the order in which a viewer reaches them. A heat map can get the broad distribution right while missing the movement between a face, a hand and an object elsewhere in the composition. The recorded scanpaths make that second challenge testable.[1]

Report an error or suggest a correction