Object Recognition and Segmentation using unstructured 3D Point Clouds

Abstract

Point Clouds are an imminent source of the scalable geometric data structure. They are widely used in computer graphics or computer graphic applications. Point clouds are in unstructured form, due to irregularities in structure the shapes are not organized in a pre-defined manner. In contrast to deep learning methods, voxelization is computationally expensive due to unnecessary voluminous data. An efficient model is required whose performance is not obstructed due to middleware dependency like in case of voxels, it is difficult to capture high-resolution features which neglect the geometric relationship between its neighbors. With the advent of deep learning and computer vision got a boost in solving problems efficiently using convolution neural networks from object recognition, and part segmentation. It is an extremely challenging task to deal with unstructured 3D point clouds using a deep neural network. To overcome this issue, we designed a new deep net architecture that directly takes raw point clouds as input. This paper proposed a novel deep architecture dubbed with edge convolution and feature transformation layers suitable for performing high-level point cloud processing tasks, including 3D object classification and part segmentation. The model works as each point of a point cloud is processed independently while maintaining geometric properties among its neighborhoods. The whole process is a recursive task where each point of the point cloud is used as an input to learn more fine-grained geometric features along with their semantic properties from potentially long distances. This novel approach achieves 1.0% better performance on 3D object recognition and part segmentation using standard datasets including ModelNet40 for object recognition and Shape Net part for part segmentation.