MARNet: Multi-scale adaptive relational network for robust point cloud completion via cross-modal fusion

Abstract

Point cloud completion, pivotal for enabling robust 3D understanding in autonomous systems and augmented reality, faces persistent challenges in structural fidelity preservation and detail adaptive restoration. This paper presents MARNet—a novel multi-scale adaptive relational network via cross-modal guidance. We first design a spatial-relational adaptive feature descriptor integrating space-feature adaptive downsampling with relation-aware weighted edge convolution. This integration effectively preserves structural integrity while suppressing outlier interference. We then introduce a hierarchical cross-modal fusion module establishing bidirectional feature interaction pathways between 3D point clouds and multi-view images through attention-guided mechanisms, which significantly enhances feature representation capacity. Additionally, our adaptive multi-resolution point generator dynamically adjusts upsampling stages based on each shape’s geometric complexity, restoring highly detailed structures while mitigating over- and under-completion issues. Extensive experiments demonstrate state-of-the-art performance with Chamfer Distance of $6.36\times 10^4$ on Completion 3D and $6.42\times 10^3$ on PCN datasets. Our method outperforms existing approaches in preserving fine details and global consistency, particularly for complex structures, while exhibiting robustness to noise and viewpoint variations. The code is available at https://github.com/long-git22/MARNet.git.