Point Enhancement Network Based on Gated Convolutional Units for 3D Shape Completion

In the field of computer vision, single-view 3D reconstruction is a challenging and ill-posed problem. Due to environmental self-occlusion, the reconstruction often suffers from noise. Our proposed approach initializes the reconstruction process with a U-Net backbone network to obtain a rough shape, which is then refined using a novel point enhancement module. This module introduces a gated convolution unit designed to enhance ambiguous point clusters within the rough shape, resulting in a more accurate and point-enhanced shape with clearer boundaries. Our proposed method also allows for the concurrent optimization of the backbone network, contributing to the field with a competitive performance and less computational overhead.