PCDM: Point Cloud Completion by Conditional Diffusion Model

During the point clouds collection process, the errors can be easily introduced into dataset, leading to biases in subsequent analyses, so the completion of point clouds is a critical step of point cloud analysis. In this paper, we introduce a conditional diffusion model architecture (PCDM) to solve point cloud completion problem. The diffusion model has shown remarkable success in image generation and has recently started to be applied in other domains, showing amazing results. By harnessing the potent generative abilities of diffusion models, we progressively get the complete point clouds derived from pure noise data. In the denosing process, we employ a Local-Global Net (LoGNet) to fuse global and local features, guiding the generation of point clouds by the model. Furthermore, to strike a balance between the completeness of the generated point cloud data and local details, we introduce an offset-attention mechanism to extract features from incomplete point clouds. Experiments conducted on multiple public datasets demonstrate that the point cloud completion method proposed in this paper outperforms previous approaches.