Point Cloud Implicit 3D Reconstruction Based on Adaptive Threshold Optimal Transport

To achieve high precision reconstruction of complex point cloud data, this paper presents a novel method based on an adaptive threshold. Firstly, a pre trained neural network is used to initially segment the point cloud, and an adaptive threshold is introduced into the discrete optimal transport to effectively detect feature regions such as edges and corners based on local geometric information. Subsequently, a zero level set smoothing loss function and a normal alignment loss function are introduced into the implicit neural network to optimize the feature region and the smooth region respectively. Finally, the optimized point cloud is reconstructed through the Marching Cubes algorithm. Experimental results show that compared with DeepSDF and Occupancy Networks, the accuracy of the proposed method on the corresponding data sets is improved by 3.33% and 1.2% respectively.