A point cloud completion network integrating Mamba and transformer architectures

Abstract

Point cloud completion aims to reconstruct complete structures from incomplete point clouds by extracting fine-grained local details and global features. Current state-of-the-art methods rely on Transformer architectures, which suffer from quadratic complexity, leading to high computational costs and trade-offs in resolution and feature extraction. To address this limitation, we propose a novel point cloud completion network that integrates the Mamba model, a state space framework with linear complexity, for feature extraction in the encoding phase. Our approach replaces the self-attention module with Mamba and introduces a multi-scale encoding network to enhance the extraction and fusion of features from incomplete point clouds. A cross-attention decoding module processes centre points and incomplete features to predict a complete point cloud. Experiments on synthetic and real-world datasets show that our method achieves comparable performance to existing state-of-the-art approaches on benchmark datasets, achieving an average CDL1 score of 6.50 on the PCN dataset. In addition, our method demonstrates superior accuracy when processing large-volume point cloud data, highlighting Mamba’s effectiveness in handling such challenges compared with Transformer-based models.