3D object Classification

Abstract

This thesis deals with the task of classifying 3D Objects. Nowadays, 3D object classification is used in robotics, self-driven cars, etc. To organize any object, we need local and global representation. As humans, we classify objects on the basis of local as well as global descriptors. Our aim is to get a better representation of a 3D object by using local and global features for task of object classification. There are several architectures through which we can get 3D object representations. Here, we use Dynamic graph convolution neural network(DGCNN) for local feature description and global representation. We use the K-means clustering algorithm in order to segment 3D objects, and every segmented cluster is represented by its centroid. We take the centroid as a global descriptor. By combining both local and global features, we try to classify the 3D object. We propose to use the k-means algorithm for obtaining global features. Specifically, centroids of clusters obtained via k-means are used as features giving a global representation of each point. We conduct several experiments to examine the quality of such features obtained for the purpose of object classification.