CM-Net: a point cloud upsampling network based on adversarial neural network

Xiaomeng Li, Chung-Ming Own, Kaizhen Wu, Qian Sun

Abstract: In the real-life application scenarios, the collected point clouds are often sparse, noisy and non-uniform, especially if the application is unable to capture the local spatial layout due to the lack of points. Thus, point cloud upsampling aims to transform sparse point clouds into dense point clouds with uniform point distribution. This study proposes a generative adversarial network-based point cloud upsampling network called CM-Net (Circular Multi-Frequency Network). The model consists of two parts: a generator and a discriminator. The generator's purpose is to convert the input sparse point cloud into a dense upsampling point cloud with our specific parts, includes multi-frequency pooling module and positive polygon-based code. On the contrary, the discriminator incudes two parts; a feature extractor and a convolutional network module, which optimizes the generator's performance. Experimental results show that our network can learn the point cloud's underlying geometric information very well, and complete the even distribution points.