Patch-Based Implicit Surface Learning Network by Combining Neighborhood Point Grouping and Fully Connected Residual Feature

Due to its lack of surface details of the underlying point cloud shape for implicit field learning, a novel patch-based surface learning network based on neighborhood point grouping and fully connected residual feature is proposed which adopts a patch-based representation for training the local sign distance field of 3D shapes. With the discrete point cloud data as input, the initial center position and radius of spherical patch representing 3D shape are obtained by applying the farthest point sampling strategy; then, the relative positions of sampling points located in each surface patch relative to its patch center can be calculated, whilst the sampling points located in the overlapping regions always exist several relative positions; for the network encoder, each relative position will be encoded into 256 dimensional latent vector using the neighboring point grouping layer, and the latent vectors of sampling points located in each surface patch are maximized to represent their latent features; for the network decoder, a multi-layer fully connected residual feature decoding module is adopted to decode the sign distance value of its relative position for each sampling point, whilst that of the sampling point located in the overlapping regions for different patches can be calculated as a weighted sum of their distance values; finally, the Marching Cube algorithm is applied to extract the zero iso-surface of sign distance values for sampling points to obtain final 3D mesh model. By using ShapeNet, ABC and Famous dataset, experimental results demonstrate that our patch-based point cloud learning network can effectively reconstruct the overall shape structure whilst maintaining its surface details. If testing on ShapeNet datasets, the IoU index is 83.9, and the CD index is 0.032. The reconstruction results by our proposed network is superior to that of DeepSDF, occupancy network, or convolutional occupancy network.