Featureless point cloud upsampling via patched-based attention

Abstract

Point clouds are a fundamental element of 3D technology, such as self-driving cars, 3D scanners, and many more. Working on it is risky, raw inputs from sensors like LiDAR present irregularities as noisy, sparse, and non-uniform. To solve this problem, point cloud upsampling is needed. For any given input, it will output a noiseless, uniform point cloud. However, generating uniform point clouds from different raw inputs is not always easy; some results present holes and some contain outliers. To address these anomalies, many approaches based on patches have been developed, such as PU-Net , PU-GAN , PU-GCN , and Grad-PU. This thesis proposes the ePUGAN model, an enhanced version of PU-GAN under implementations of new advanced deep learning tools such as MAMBA, P3DNet, and recently the midpoint interpolation from arbitrary-scale point cloud upsampling. The goal is to study the behavior of ePUGAN trained on the PU1K dataset based on the new component implementation within the overall networks. After consideration, the presence of MAMBA in one of the adversarial blocks outperforms the original PU-GAN, and the implementation of the arbitrary-scaling also during training improves the performance of ePUGAN , but putting P3DNet as a feature extractor does not give a total guarantee of quality. Lastly, for high performance and stability, it is highly recommended to use ePUGAN trained on arbitrary-scaling with or without MAMBA.