HMM-based Convolutional LSTM for Visual Scanpath Prediction
Ashish Kumar Verma, Debashis Sen · 2019
The human visual system performs a dynamic process of scanning the scene by rapid eye movements and fixations, yielding a visual scanpath. We propose an approach to generate artificial visual scanpaths for natural images. A convolutional long short term memory (LSTM) neural network is employed, which learns the mapping of image features to eye fixations by modeling the sequential dependencies of the fixations in a scanpath. A novel approach of hidden Markov model (HMM) based data augmentation is presented that increases the number of available image-specific input-output pairs to train the LSTM appropriately. Both the HMM and the LSTM are designed to be consistent with existing knowledge on saccadic eye movements. Experimental results on a standard eye-tracking dataset demonstrate that the proposed approach does better than the state-of-the-art and generates realistic visual scanpath data.