Fig 1.
Experimental scanpath and fixations density.
Left. An image and a scan path. Each dot is a fixation and the dashed line illustrates the saccade. Right. The empirical fixation density map as generated by aggregating the fixations from all subjects for a given image.
Fig 2.
Local and global attention maps.
On the left is an example of an empirical saliency map, the dot indicates a fixation location. On the right are the probability maps generated by either the local attention (upper panel) or the global attention policy (lower panel). The arrow indicates a saccade.
Fig 3.
Experimental and simulated scan paths.
Left. An image and a scan path recorded from a human observer. Right. The experimental static saliency map and a scan path generated by the Local and Global Attention Model. The green arrow pointing right represents the second randomly selected fixation location z2. The green arrow pointing left represents the last fixation in the scan path. The blue dots are fixations that were generated from a global attention step and the pink dots are fixations that were generated from a local attention step. The experimental data shows clearly the phenomenon of saccadic momentum which is not captured by the model. This is further discussed in the Results and Discussion.
Fig 4.
Inference results on simulated data.
Model parameter recovery. To test the inference algorithm we fit the model to simulated data with known parameters values. Each panel includes the inferred posterior distribution of each parameter after the inference process. The ten curves present 10 different inference processes starting from different values. The vertical lines are the values with which the data was generated. The black dashed curve is the prior distribution. The plotted densities are not normalized.
Fig 5.
Comparison between the empirical saliency map and the fixation density of data generated by the model, for three different images from the test-set.
The empirical saliency is represented by the shading, and the contour lines represent the density of the data generated by the model. The generated data recovers the original empirical saliency map.
Fig 6.
Saccade amplitude density—Experimental and simulated.
Saccade amplitude density, aggregated over the data from all participants, of the experimental data and data generated by the full model and the simplified competitor models. Top. Comparison of all models. Bottom. Comparison between the full model and the Local Saliency Model. The shading corresponds to confidence bounds regarding the estimate of the model parameters. The full model captures the different kinds of saccade lengths, whereas the simpler models fail to do so.
Fig 7.
Saccade amplitude autocorrelation—Experimental and simulated.
Saccade amplitude autocorrelation, averaged over experimental data from all participants and over all simulations generated by the full model (and the various competitor models). The full Local and Global Attention Model approximates the autocorrelation in amplitude of successive saccades, whereas the simpler models fail to reproduce the lag-1 anti-correlation.
Fig 8.
Comparison between experimental and simulated subjects’ mean and standard deviation of saccade amplitude.
Left. Participants’ mean saccade amplitudes compared with the mean saccade lengths of data generated by the Local and Global Attention model. Right. The standard deviation of the subjects’ saccade amplitude compared to the standard deviation of the data generated by the Local and Global Attention model. Overall the model captures the both the mean and the standard deviation of the saccade amplitude of the different subjects.
Table 1.
Comparison of the coefficient of determination, between the mean (or std) of the subjects’ saccade amplitudes and the saccade amplitudes of the data generated by the different models.
Other than the local saliency model, all models capture both the mean ant the standard deviation of the saccade amplitudes of the different subjects.
Fig 9.
Experimental and simulated saccade direction and saccade direction change frequency.
Left. Absolute saccade direction. The empirical data demonstrate a strong tendency to saccades towards the left and right directions, and a weaker tendency to perform saccades directed upwards. The generated data captures the tendency to perform horizontal saccades, but not vertical saccades. Right. Change in saccade direction. The generated data demonstrates the tendency to persist in the same direction. The models fail to capture this persistence.
Fig 10.
ROC curves of the different model variants.
The full model performs slightly better than the Local and Fixed Choice models. The Local Saliency Model performs significantly worse than the other model variants.
Table 2.
AUC, NSS and IG of the different model variants.
The full model performs better than the Local and Fixed Choice models. The Local Saliency Model performs significantly worse than the other model variants.