Research On Object Proposal Algorithm For RGB-D Indoor Scenes

Object proposal,as the key step of scene understanding,is one of the hotspots in the field of machine vision.When applying most existing object proposal methods on complex indoor scenes,the results show that there are some problems such as ignoring the small size object or objects in planar regions and detection inaccuracies caused by occlusion or similar color distributions in foreground and background areas.In view of these problems,this paper carries out the research of using depth information to reduce the error of object proposals in indoor scenes,and completed the following three aspects of work:1.In this paper,an estimation method of scene layout based on depth information is proposed.The algorithm uses weighted bilateral filtering with neighborhood pixel depth value to fill the holes in the depth map and translates the filled depth map into point cloud data.The RANSAC algorithm is used for plane segmentation and classification considering the three-dimensional Euclidean distance between the spatial points.The simulation experiments show that the algorithm uses the 3D features to effectively reduce the big object proposal error caused by the occlusion.2.Aiming at the problem that small size objects and objects in the planar area are easy to be ignored,a region hierarchical clustering algorithm based on image segmentation is proposed.The graph based segmentation and constrained parametric min-cuts algorithm are applied to get multi-scale superpixels after preprocessing the RGB-D image.Then four similarity measures are utilized for region hierarchical grouping to obtain all scales of candidate regions.The experiment shows that the results of this algorithm effectively improve bounding box proposal recall score with fewer object candidates.In addition,the algorithm does not need pre-training process.3.In this paper,an instance segmentation algorithm based on improved GrabCut is proposed to solve the problem that the segmentation effect of GrabCut is not ideal when the foreground and background have similar color distributions.With object bounding boxes as input,the algorithm models foreground and background pixels with six dimensional Gauss mixture models.Then a Gibbs energy formula with a graph representation are established according to pixels in the bounding box and obtain the object instance segmentation using min-cut algorithm.The simulation experiment shows that compared with state-of-the-art methods,the algorithm obtains higher Jaccard index and can obtain more accurate result of object instance segmentation.