In this paper we present an approach for detection of simple objects in RGB-D data. Object detection in cluttered indoors environments is an important perceptual capability of robotic systems required for object search and pick and deliver tasks. For long term autonomy robots should learn how objects look like and where they appear in an weakly supervised manner. In this work we exploit the depth information to provide evidence about occlusion boundaries and scale of the objects. The depth discontinuities along with image contours computed in the vicinity of the detection window boundary form an (em objectness) measure, which is used to train an SVM classifier. In the testing stage we exploit the knowledge of the actual size of the object to propose the scale of the detection window significantly pruning the number window candidates to be evaluated. We evaluate our approach for detecting simple objects on NYU RGB-D dataset, illustrate the effectiveness of our approach as well as difficulties with the standard evaluation methodologies.
展开▼
