CFNet: Point Cloud Upsampling via Cascaded Feedback Network

Xuan Wang, Yi Li, Linna Wang, Li Lu

Abstract: The rapid development of point cloud processing has ushered in a new era of point cloud upsampling. However, most existing methods for point cloud upsampling focus on designing feed-forward cascaded networks based on a coarse-to-fine pipeline to enhance the network’s performance. Unfortunately, these methods overlook the potential benefits of incorporating higher-level information to improve low-level feature learning. To address this issue, we propose a novel architecture called Cascaded Feedback Network (CFNet), which differs from previous methods by incorporating both feed-forward and feedback mechanisms. The feedback mechanism in our CFNet can enhance the feature learning of the low-level layer by fusing the information from the high-level layer. Additionally, we propose a novel Feedback Upsampling (FU) module to construct our CFNet. Through extensive experiments on synthetic datasets such as PU1K and PU-GAN datasets, we demonstrate that our proposed CFNet architecture, along with the FU module, outperforms existing methods in point cloud upsampling, indicating the effectiveness of our proposed approach.