Potsdam data set of Eye Movement On Natural Scenes (DAEMONS)

Abstract

Modeling eye movement behavior during natural scene viewing is an important approach in understanding visual cognition. As static location prediction has reached very high standard, scientific interest has expanded to modeling gaze trajectories over time. In order to accommodate a variety of different modeling approaches, ranging from Deep Neural Networks to biologically inspired mechanistic models, research in this field requires high-quality, openly available data sets for model training and benchmarking. Existing data sets that have been used for saliency prediction typically lack comprehensive temporal information for modeling fixation sequences, or else fall short in terms of size (numbers of images and participants) or data quality (such as spatial and temporal resolution). To address this gap, we collected and published a purpose-built scene viewing corpus data set, comprising 984,000 fixations on 2,400 unique images with 250 participants, with explicit training-, test-, and validation subsets. This data set is intended to serve as a benchmark for various modeling approaches and facilitates cross-disciplinary research, enabling comparisons between machine learning techniques and insights theoretical modeling. The availability of this comprehensive data set enhances our understanding of eye movement behavior and promotes advancements in the field.

Citation

Currently under review at Frontiers Psychology.

Log of known issues

12.4.2024 - Two images slipped through our rigorous process of image selection: in the Potsdam image dataset image 493 and 681 show the same photograph. The same is true of images 831 and 832.