PROJECT TITLE :

RGB-D Saliency Detection With Pseudo Depth

ABSTRACT:

The use of depth information in the recognition of salient objects has been demonstrated in recent studies. There are still a lot of photos out there that don't include any depth information. A 3D view of a scene can be obtained from an RGB-only image by the human brain, which is capable of extracting the geometric model of the scene. We are inspired by this observation to present a novel idea, called RGB-'D' saliency detection, which extracts pseudo depth from RGB images and then performs 3D saliency identification based on this information. To improve the performance of classic RGB saliency models, pseudo depth can be used as picture characteristics, previous knowledge, an additional image channel, or independent depth-induced models. With the help of the pseudo depth, we have devised an example of a new salient object recognition method that leverages it to build an algorithmic prior and a depth contrast feature. Promising results have been demonstrated through extensive testing on a variety of standard databases. Our RGB-'D' saliency framework is also enhanced by the addition of two supervised RGB saliency models. The results show that the proposed RGB-'D' saliency paradigm can be generalised.