Point Cloud Completion Network Based on Dynamic Feature Selection

In response to the issues of incomplete feature extraction and point cloud shape deformation existing in the current point cloud completion process, this paper proposes a dynamic feature selection point cloud completion network. The network adopts an encoder-decoder structure, and a gate-controlled feature matching module is designed to integrate the feature sequence throughout the network, enabling the point cloud network to extract more comprehensive features. In the process of decoding and generating point clouds, the gate-controlled feature matching module is combined with the attention mechanism, and a hierarchical feature recovery strategy is adopted to restore the complete point cloud. The experimental results on the public dataset ShapeNet show that the proposed network model achieves better completion results in the point cloud completion task, further improving the point cloud completion quality.