Comparison of strategies for scan-path prediction using CNN-based saliency maps

Thumbnail Image
Publication date
Defense date
Journal Title
Journal ISSN
Volume Title
Google Scholar
Research Projects
Organizational Units
Journal Issue
any different approaches exist to understand where people’s attention is concentrated. However, very few information concerning the order in which these regions should be scanned is given. Developing a system that allows obtaining sequences of fixations automatically given an image is the main goal addressed in this work. In order to do that, a CNN-based model is implemented to generate saliency maps from the SALICON dataset. These estimated saliency maps are analyzed using NSS, KL and shuffled AUC metrics in order to evaluate its performance and ensure its usability for later generating the scan-paths from them. The scan-paths are sampled using three different sampling strategies, which are analyzed and compared using the Jarodzka metric, after validating their corresponding parameters. Our experiments show that realistic scan-paths are possible to obtain. However, the accuracy and precision of these sequences of fixations depend on the sampling strategy used. The strategy with the best results corresponds to a function that calculates local maxima given a distance. Results also show that the distance between fixations should be equal to the third of the image size, resulting in scan-paths with small lengths.
Convolutional Neural Network (CNN), Complex system, Biological neurons, Artificial neurons, Saliency maps, Computer vision, Image processing
Bibliographic citation