Temporal distribution of saccades with deep learning salience maps
The classic Saliency Model by Itti and Koch launched many studies that contributed to the modelling of layers for vision and visual attention. The aim of this study is to improve the existing saliency model by using a neural network to generate salience maps to model human saccade generation. The proposed model uses a Leaky Integrate-and-Fire layer for temporal predictions, and replaces spatial salience with a deep learning neural network in order to create a generative model that combines spatial and temporal predictions. The results involve a deep neural network which is able to predict eye movements based on unsupervised learning from raw image input, as well as supervised learning from fixation maps retrieved during an eye-tracking experiment with 35 participants at later stages in order to train a 2D softmax layer. The results imply that it is possible to match model human fixation locations but temporal distributions are still limited by the accuracy of the leaky algorithm.