Progressive Framework of Learning 3D Object Classes and Orientations from Deep Point Cloud Representation

Abstract

Deep learning approaches to the classification and orientation estimation of 3D objects have not been so successful to date. The increased complexity of dealing with 3 axes orientation variations associated with individual objects is to be blamed. To date, the existing approaches have shown limitations in their capacity of handling orientation variations in trade-off with their performance. This paper presents a progressive framework of learning 3D objects in terms of their classes and 3D orientations that overcomes the above complexity induced trade-off. By a progressive framework, we mean that the object classification and the estimation of three axes of orientations are learned one after another progressively based on whatever learned previously as prior knowledge. The proposed framework, referred to here as 3D POCO Net, is configured with multiple point cloud based deep networks that are cascaded through the association of their learned global features. As a modular architecture, 3D POCO Net offers not only efficiency and generality in representation and training but also expandability due to the progressive nature of learning. The proposed 3D POCO Net is implemented for full 3 axes orientation variations and trained with about 2.4 million orientation variations generated from ModelNet10. The high accuracy in object classification and orientation estimation verified experimentally for a large scale of 3 axes orientation variations indicates that the proposed progressive learning approach is able to overcome the aforementioned complexity induced trade-off.