3D Saliency Based on Supervoxels Rarity in Point Clouds

Visual saliency is a computational process that seeks to identify the most attention-drawing regions from a visual point of view. Most methods of salience are based on characteristics such as color, texture and more recently tried to introduce the depth information, which is known to be an important cue in the human cognitive system. We present a new full 3D mechanism of computational attention for extracting salient information directly from a 3D point cloud which can be either on a single RGBD view or the fusion of multiple viewpoints. The proposed method reduces the point cloud complexity through over-segmentation in order not to process all voxels information, but only supervoxels, which are larger color an geometrically consistent 3D structures. A very interesting feature of our algorithm is that it provides a bottom-up saliency map which is viewpoint independent. This bottom-up saliency map can then be specialized by using topdown information for a given viewpoint by adding a centered Gaussian and the viewpoint-dependent depth information. Our approach was tested on a complex database of 80 color depth and picture pairs with associated pixel level ground truth segmentation masks. The use of 3D Point Clouds improves the results compared to depth maps extended models even if only the color feature is used.