Disentangling top-down vs. bottom-up and low-level vs. high-level influences on eye movements over time

Top-down and bottom-up, as well as low-level and high-level factors influence where we fixate when viewing natural scenes. However, the importance of each of these factors and how they interact remains a matter of debate. Here, we disentangle these factors by analysing their influence over time. For this purpose we develop a simple saliency model, which uses the internal representation of a recent early spatial vision model as input and thus measures the influence of the low-level bottom-up factor. To measure the influence of high-level bottom-up features, we use DeepGaze II, a recent DNN-based saliency model. To vary top-down influences, we evaluate the models on a large Corpus dataset with a memorization task and on a large visual search dataset. We can separate the exploration of a natural scene into three phases: The first saccade, which follows a different narrower distribution, an initial guided exploration characterized by a gradual broadening of the fixation density and an equilibrium state which is reached after roughly 10 fixations. The initial exploration and the equilibrium state target similar areas which are determined largely top-down in the search dataset and are better predicted on the Corpus dataset when including high-level features. In contrast, the first fixation targets a different fixation density and contains a strong central fixation bias, but also the strongest guidance by the image. Nonetheless, the first fixations are better predicted by including high-level information as early as 200ms after image onset. We conclude that any low-level bottom-up factors are restricted to the first saccade, possibly caused by the image onset. Later eye movements are better explained by including high-level features, but this high-level bottom-up control can be overruled by top-down influences.