Object Recognition and Segmentation in Indoor Scenes from RGB-D Images

We study the problem of automatic recognition and segmentation of objects in indoor RGB-D scenes. We propose to formulate the object recognition and segmentation in RGBD data as a binary object-background segmentation, using an informative set of features and grouping cues for small regular superpixels. The main novelty of the proposed approach is the exploitation of the informative depth channel features which indicate presence of depth boundaries, the use of efficient supervised object specific binary segmentation and effective hard negative mining exploiting the object co-occurrence statistics. The binary segmentation is meaningful in the context of robotics applications, where often only an object of interest needs to be sought. This yields an efficient and flexible method, which can be easily extended to additional object categories. We report the performance of the approach on NYU-V2 indoor dataset and demonstrate improvement in the global and average accuracy compared to the state of the art methods.