Eye movements on natural videos: low-level feature correlations at successive fixations

Michael Dorr, Karl Gegenfurtner, and Erhardt Barth

We investigated the relationship between low-level image features at successive fixations during free-viewing of natural dynamic scenes.

We recorded 37000 fixations from 54 subjects watching 18 high-resolution video clips of outdoor scenes. Colour, local orientation, and motion were computed on a Gaussian multiresolution pyramid (6 spatial, 3 temporal levels). We also computed geometrical invariants that have been used to predict eye movements before.

We then examined the distributions of differences between features at successive fixations, i.e. their temporal correlations along the scanpath. We compared the recorded scanpaths with randomly  generated scanpaths with varying degree of similarity to natural scanpaths (saccade length, direction, etc.).

In most conditions, we found statistically significant differences in the distributions. However, we found that these feature correlations along the scanpath are mainly due to spatio-temporal correlations in natural scenes and a bias in target selection (e.g., moving objects are fixated more often).

Therefore, we conclude that low-level features at the current centre of fixation contribute little to the selection of the next saccade target.