A Deep Model of Visual Attention for Saliency Detection on 3D Objects

A variety of saliency detection techniques have been proposed during the last two decades to determine important regions on the surface of 3D shapes in form of triangular meshes. However, most fail in predicting the regions where human eyes naturally fixate when observing and exploring an object. Taking inspiration from biological studies that enumerate a list of object characteristics revealed in human visual processing and the influence of semantic properties in the emergence of neural responses in human brain, in this work, we propose a deep convolutional neural network architecture using gradient-based class activation mapping to detect saliencies on the surface of 3D objects when classifying them based on their different properties. We further argue that using Pearson Correlation Coefficient is not sufficient for the evaluation of saliency values and therefore propose a novel evaluation technique to determine how reliable is the detection performed by saliency detectors to predict eye fixations. More specifically, this evaluation metric measures the distance between the most salient region detected and the respective location of human eye fixation. Evaluating the results based on visual comparison, as well as using the proposed evaluation technique, demonstrates that our model is successful in predicting the locations where human eye fixates. Results are compared with five state-of-the-art saliency detectors, and our experiments suggest that in average the location of the highest saliency detected by our approach is closer to the location of human eye fixation by about 22.55% to 77.76% in comparison with five state-of-the art method" on a public dataset.