Gaussian point cloud autoencoder for partial point cloud reconstruction

Partial point cloud reconstruction supports 3D perception under occlusion and limited sensing. We propose the Gaussian Point cloud AutoEncoder (GPAE), a lightweight network that reconstructs a complete point cloud from one partial view. A graph encoder infers explicit Gaussian (means and variances), and patch decoders for generating local surfaces. Weight vectorization and decomposition reduce Gaussian encoder parameters by at least 87.97 % , while random viewpoint sampling improves sparse input generalization. On ShapeNetCore, GPAE-16 achieves an average dense output Chamfer Distance of 0.56 × 10 − 3 at N 3 = 8192 with 1.55 × 10 6 parameters, lower than those of the matched resolution AtlasNet, PCN, and MSN baselines. At 50 % point reduction, average CD increases by 8 % . The explicit components expose local position and spatial extent.