Latent Feature-Aware and Local Structure-Preserving Network for 3D Completion from a Single Depth View

Abstract

Recovering the geometry of an object from a single depth image is an interesting yet challenging problem. While the recently proposed learning based approaches have demonstrated promising performance, they tend to produce unfaithful and incomplete 3D shape. In this paper, we propose Latent Feature-Aware and Local Structure-Preserving Network (LALP-Net) for completing the full 3D shape from a single depth view of an object, which consists of a generator and a discriminator. In the generator, we introduce Latent Feature-Aware (LFA) to learn a latent representation from the encoded input for a decoder generating the accurate and complete 3D shape. LFA can be taken as a plug-and-play component to upgrade existing networks. In the discriminator, we combine a Local Structure Preservation (LSP) module regarding visible regions and a Global Structure Prediction (GSP) module regarding entire regions for faithful reconstruction. Experimental results on both synthetic and real-world datasets show that our LALP-Net outperforms the state-of-the-art methods by a large margin.