Abstract Existing computational models of visual attention generally employ simple image features such as color, intensity or orientation to generate a saliency map which highlights the image parts that attract human attention. Interestingly, most of these models do not process any depth information and operate only on standard two-dimensional RGB images. On the other hand, depth processing through stereo vision is a key characteristics of the human visual system. In line with this observation, in this study, we propose to extend two state-of-the-art static saliency models that depend on region covariances to process additional depth information available in RGB-D images. We evaluate our proposed models on NUS-3D benchmark dataset by taking into account different evaluation metrics. Our results reveal that using the additional depth information improves the saliency prediction in a statistically significant manner, giving more accurate saliency maps.

BibTeX