Abstract
Recent research from our lab has shown that gaze patterns of subjects viewing color photographs are predictive of scene category. Here we ask if this result extends to grayscale photographs and line drawings. 77 participants viewed grayscale photographs and line drawings of real-world scenes. In a leave-one-subject-out cross validation analysis, viewed scene category was predicted from gaze patterns by computing the best match between the gaze path of the left-out subject in a given trial and a set of fixation density maps (FDMs) that were derived from all other subjects. Gaze patterns are predictive of scene category for photographs (accuracy=33.6%, significantly above chance level of 16.7% with p=3.1·10-52) and line drawings (accuracy=30.9%, p=4.5·10-43). Predicting category of line drawings from FDMs of photographs was accurate for 28.5% (p=6.6·10-39) of the trials, and predicting photographs from FDMs of line drawings for 29.9% (p=5.0·10-50). However, prediction accuracy across image types was significantly lower than accuracy within image type (p=4.5·10-11). This pattern of results suggests that gaze patterns between line drawings and photographs are compatible to a limited extent. We investigated the temporal aspect of gaze patterns by restricting the analysis to time bins with a width of 300 ms. In all four cases, prediction accuracy increased for the first few hundred milliseconds, peaked at 600ms after stimulus onset, and then slowly decreased. What are the image properties that drive gaze behavior that is so distinctive of scene category? We have recently shown that localized features, such as contour junctions, are important for scene categorization. Using the line drawings, we are currently investigating whether subjects look at those important features and use them to guide their gaze patterns.
Meeting abstract presented at VSS 2015