Attention-based structural analysis point cloud completion network

Abstract

Point cloud completion involves repairing incomplete and fragmented point clouds into complete ones to facilitate downstream tasks more effectively. In current learning-based methods, if the network fails to adequately consider the structural relationships between point cloud features and the prior knowledge embedded in incomplete point clouds during learning, it may lead to completed point clouds lacking in local details, thereby affecting the completion accuracy. To address this issue, this paper proposes an improved network, namely the Attention-based Structural Analysis Point Cloud Completion Network. The entire network continues the structure of the pf-net network, employing a cascaded approach for feature extraction to extract and merge high-dimensional features. It introduces a self-attention mechanism to analyze the structural relationships inherent in high-dimensional features, and a cross-attention mechanism to assist the decoder in utilizing the prior features from incomplete point clouds to recover complete point clouds. Experimental results on the Shapenet dataset show a 4.818% improvement in completion accuracy compared to the baseline network. Additionally, there are better visual results, and the network demonstrates advantages in CD compared to other networks.