MCNet: Multi-3D point cloud completions with diverse latent shape prior
Authors: Xi, L., Tang, W., Wan, T.R., Xue, T.
Journal: Knowledge Based Systems
Publication Date: 22/04/2026
Volume: 339
ISSN: 0950-7051
DOI: 10.1016/j.knosys.2026.115574
Abstract:A 3D point cloud completion task is to generate one or multiple completed 3D objects based on incomplete and partial input data. While single-output models suffer from the large missing input data, current multiple-output models require texts, images or depth maps to guide completion tasks, dramatically increasing network model sizes with limited output diversities. This paper proposes a novel probabilistic architecture, MCNet, to address these challenges. MCNet includes a latent shape prior module and a conditional probabilistic diffusion module to explore multiple completion results faithful to the partial input data. We hypothesise that the learnt latent space contains diverse latent shapes. Therefore, through a reverse process of latent shape prior distribution, we can recover a diverse range of latent shapes from the random samples of a Gaussian distribution. Based on the diverse latent shapes and the latent shape of the partial input, we design a conditional probabilistic diffusion model to convert a noise distribution into the distribution of multiple 3D shapes. Critically, MCNet is lightweight and generates flexible output point cloud resolutions without any further training while maintaining the model size constant with the increase of the input and output point cloud resolutions. Experimental results demonstrate that MCNet achieves state-of-the-art performance and also exhibits strong generalization to unseen 3D objects and real-world 3D scenes that are never trained. Code is available at https://github.com/LONG-XI/MCNet.
Source: Scopus