Improved Scan Path Model Using Visual Saliency Dataset

Abstract

Based on PathGAN network, this paper improves it in many aspects, and obtains a more excellent deep neural network for detecting scan path. Scan path is a sequence of fixation points generated by human eyes on stimuli such as images or videos in scene exploration tasks. In this paper, we use the line of sight landing point data on the visual saliency dataset mit1003 to convert it into scan path. Then, the scan path is used in the improved PathGAN for training, and the scan path generated by the model is more diverse. At the same time, the new scan path representation method is used to increase the information of scan path, which can be used to predict the scan path of stimulating pictures in other datasets, and expand the scope of application of the model.