A method to create real-like point clouds for 3D object classification

Abstract

There are a large number of publicly available datasets of 3D data, they generally suffer from some drawbacks, such as small number of data samples, and class imbalance. Data augmentation is a set of techniques that aim to increase the size of datasets and solve such defects, and hence to overcome the problem of overfitting when training a classifier. In this paper, we propose a method to create new synthesized data by converting complete meshes into occluded 3D point clouds similar to those in real-world datasets. The proposed method involves two main steps, the first one is hidden surface removal (HSR), where the occluded parts of objects surfaces from the viewpoint of a camera are deleted. A low-complexity method has been proposed to implement HSR based on occupancy grids. The second step is a random sampling of the detected visible surfaces. The proposed two-step method is applied to a subset of ModelNet40 dataset to create a new dataset, which is then used to train and test three different deep-learning classifiers (VoxNet, PointNet, and 3DmFV). We studied classifiers performance as a function of the camera elevation angle. We also conducted another experiment to show how the newly generated data samples can improve the classification performance when they are combined with the original data during training process. Simulation results show that the proposed method enables us to create a large number of new data samples with a small size needed for storage. Results also show that the performance of classifiers is highly dependent on the elevation angle of the camera. In addition, there may exist some angles where performance degrades significantly. Furthermore, data augmentation using our created data improves the performance of classifiers not only when they are tested on the original data, but also on real data.