Hierarchical Dilated Dense Graph Convolution Network for Point Cloud Upsampling

Point cloud upsampling technology has been studied in detail to address the sparsity and non-uniform distribution problems commonly found in real point cloud data. The upsampling methods based on graph convolutional networks have achieved superior performance, but it still has shortcomings in multi-scale feature fusion and information transmission efficiency. To this end, we propose a new point cloud upsampling method based on graph convolution, PU-HDDGCN. The core of this method is to construct a Hierarchical Dilated Dense Graph Convolutional Network(HDDGCN) as a feature extractor. Through the collaborative feature extraction of multi-level receptive fields, multi-scale fusion of geometric topological information is achieved, effectively overcoming the limitations of a single-scale receptive field. In addition, we propose using the attention mechanism to improve the periodic shuffle upsampling module, the module named Attention Periodic Shuffle(APS), effectively enhances the smoothness of the generated point set through a dynamic feature weighting mechanism. Experimental results show that this method significantly outperforms existing methods in terms of geometric fidelity, while achieving more efficient inference performance with fewer parameters.