Abstract
The distributions of saccades and fixations during natural scene viewing are influenced by image independent biases [Tatler & Vincent 2008, J. Eye Movement Research]. One example is the central bias [Clarke & Tatler 2014, Vis. Res.], which accounts for a substantial component of fixation placement when viewing static images. We introduce the concept of saccadic flow, a generalisation of the central bias that describes the image-independent conditional probability of making a saccade to (xi+1, yi+1), given a fixation at (xi, yi). This allows us to capture the distribution of saccadic amplitudes, directions, along with fixation distributions. The model is fitted to data from eight previously published datasets of eye movements and takes the form of a multivariate Gaussian distribution of saccade landing positions (xi+1, yi+1), for a fixed starting position, (xi, yi). The parameters of the Gaussian distribution are taken to be polynomial functions of (xi, yi) and are modelled using robust regression. We then compute the likelihood of different datasets (the eight training sets, along with five previously help out test sets), and compare to the likelihood of the data given the central bias and a uniform baseline. In all cases, saccadic flow offers a large improvement over the two other baselines. We suggest that saccadic flow can be used as a useful prior when carrying out analysis of fixation locations, and can be used as a sub-module in models of eye movements during scene viewing. We demonstrate the utility of this idea by presenting bias-weighted gaze landscapes, and show that there is a link between the likelihood of a saccade under the flow model, and the salience of the following fixation.
Meeting abstract presented at VSS 2017