Comparing saliency maps and eye-tracking focus maps: The potential use in visual impact assessment based on landscape photographs

• Published in: Landscape and urban planning 2016-04, Vol.148, p.17-26
• Main Authors: Dupont, Lien, Ooms, Kristien, Antrop, Marc, Van Eetvelde, Veerle
• Format: Article
• Language: English
• Subjects: Assessments; buildings; Correlation; Correlation analysis; eyes; Human; humans; Landscape design; Landscapes; planning; prediction; urban areas; Urbanisation; urbanization; Viewing; Viewing pattern; Visual; Visual conspicuity; Visual landscape integration; visual perception
• ISSN: 0169-2046; 1872-6062
• DOI: 10.1016/j.landurbplan.2015.12.007

•Saliency maps are reliable predictions of the human visual attention distribution.•Non-salient objects provide an optimal visual integration into the landscape.•Saliency maps can also be used to identify high visual impact designs (landmarks). In this study, we analyse how well saliency maps, which are theoretical predictions of the human viewing pattern, are correlated with human focus maps, obtained by tracking 42 observer's eyes while free-viewing landscape photographs ranging from rural to urban environments. The Pearson's correlation coefficient was calculated on the predicted and measured pixels’ greyscale values. A relatively high correlation was obtained, indicating that the saliency maps can be used as reliable predictions of the human observation pattern and thus can predict which elements in a landscape will catch the attention. These findings are useful in visual impact assessment, a step in the planning process which is often not well elaborated or even skipped. Saliency maps could, for instance, be used to compare the conspicuity of different designs of a construction when simulated in photographs of the original landscape. As the visual impact of an object is reduced when its visual perception decreases, the least salient design will also have the lowest visual impact and will correspond to the best integration into the existing landscape. This method is easy and produces an objective measure of the degree of visual impact. However, as slight differences in correlation depending on the degree of urbanisation of the landscape were found, this methodology will not be equally reliable in all types of landscapes. Predictions of the viewing pattern in rural landscapes with a limited amount of buildings have been demonstrated to be most reliable. In more urbanised landscapes this reliability slightly decreases but nevertheless remains significant.