GridFix: A Python toolbox to facilitate fixation analysis and evaluation of saliency algorithms using Generalized linear mixed models (GLMM)

Why do humans fixate some areas of a scene and not others? Typically, this question is addressed by comparing features at fixated locations to baseline locations. Baseline choice is critical and often needs post-hoc sampling of fixated locations from other scenes, observers and/or conditions. To overcome this issue, we proposed to combine a-priori parcellation of scenes with GLMM (Nuthmann & Einhäuser, 2015). This approach also automatically accounts for feature dependencies and quantifies the unique contribution of individual features and other predictors, such as spatial biases. Here we present an open-source Python implementation of this method. The toolbox computes standard image features, allows for importing custom-made feature and salience maps, and provides scene parcellations of arbitrary granularity. It outputs a predictor matrix and source code to run GLMMs with lme4 in R, which the user can then adapt further. To exemplify the typical workflow, we compare different salience algorithms and how well they predict fixations above and beyond the central bias. Recent models outperform the classical saliency model of Itti et al. (1998). However, the effect of central bias varies between algorithms, suggesting that some augment their performance by (explicitly or implicitly) incorporating the tendency to look at scene centres.