Predicting Saliency Maps for ASD People

This paper presents a novel saliency prediction model for children with autism spectrum disorder (ASD). We design a new convolution neural network and train it with a new ASD dataset. Among the contributions, we can cite the coarse-to-fine architecture as well as the loss function which embeds a regularization term. We also discuss about some data augmentation methods for ASD dataset. Experimental results show that the proposed model performs better than 6 models, one supervised model finetuned with the ASD dataset. Contrary to control people, our results hint that no center bias apply in visuall attention for autistic children.