3D object categorization in cluttered scene using deep belief network architectures

Abstract

3D object classification in cluttered scenes is a critical area of computer vision and robotic research for autonomous robots to act in their surrounding area. In this chapter, we extend our previous work [51] by classifying 3D object categories in real-world scenes. We extract geometric features from 3D point clouds using a 3D global descriptor called Viewpoint Feature Histogram (VFH) then we learn the extracted features with Deep Belief Networks (DBNs). Thereafter, we test the power of Discriminative and Generative DBN architectures (DDBN/GDBN) for object categorization. The experiments on Washington RGBD dataset demonstrate the robustness of discriminative architecture which outperforms state-of-the-art. Also, we evaluate the performance of our approach on the real-world objects that are segmented from cluttered indoor scenes.