3D Object Categorization and Recognition based on Deep Belief Networks and Point Clouds

Abstract

3D object recognition and categorization are an important problem in computer vision field. Indeed, this is an area that allows many applications in diverse real problems as robotics, aerospace, automotive industry and food industry. Our contribution focuses on real 3D object recognition and categorization using the Deep Belief Networks method (DBN). We extract descriptors from cloud keypoints, then we train the resulting vectors with DBN. We evaluate the performance of this contribution on two datasets, Washington RGB-D object dataset and our own real 3D object dataset. The second one is built from real objects, following the same acquisition conditions than those used for Washington dataset acquisition. By this proposed approach, a DBN could be designed to treat the high-level features for real 3D object recognition and categorization. The experiment results on standard dataset show that our method outperforms the state-of-the-art used in the 3D object recognition and categorization.