This code allows you to choose a blur sigma and a center weight for your saliency maps, by optimizing these parameters on a separate dataset (a subset of the MIT ICCV saliency dataset). This code implements the optimization as a simple grid search over blur and center parameters, and then chooses the best setting under the ROC metric on the ICCV data. Please feel free to write your own optimization code. We also provide code for histogram matching saliency maps to a human fixation map distribution, as this tends to increase performance on some of the metrics. However, performance benefits according to all these optimizations are model-dependent, and the ICCV data provides a good testbed for trying them out and measuring their effects. The ICCV data is an appropriate choice because a very similar eyetracking set-up and data collection procedure was used for this data as for the benchmark dataset.
Latest commit f69abe0
Sep 25, 2014
|Failed to load latest commit information.|