An Enhanced Generative Adversarial Network for 3D Point Cloud Upsampling

We present an enhanced generative adversarial network for 3D point cloud upsampling (PU-EGAN). The Residual Dense Feature Extractor (RDFE) improves the network’s ability to capture subtle features and enhance feature representation. It achieves this by combining residual concatenation with dense feature propagation. To further improve the reconstruction of point cloud details and textures, we propose Point Attention Coordinate Reconstruction (PACR). PACR focuses on generating more detailed points by calculating point attention weights using the newly introduced Point Attention (PA) mechanism. Both objective and subjective evaluations demonstrate that PU-EGAN outperforms existing methods.