Attention Based Multi-Scale Feature Fusion for Point Cloud Completion

Point cloud completion is a crucial task aimed at generating a high-resolution and complete 3D representation from sparse or incomplete observations. In this research, we introduce an innovative attention-based multi-scale feature fusion network (AMF-Net) to tackle the problem of recovering geometric information at the finer details in point cloud completion. AMF-Net adopts a deep learning framework with an encoder-decoder architecture, comprising three key components: a feature extraction module, a seed generation module, and a point generation module. These components operate in a coarse-to-fine manner to efficiently reconstruct the point cloud. An attention module is incorporated during the feature extraction phase to effectively capture both global and local geometric features, eliminating need for explicit region division methods such as K-Nearest Neighbors (KNN). Integrating the attention mechanism with multi-scale feature fusion enables AMF-Net to process large-scale point cloud data more effectively, optimizing both performance and computational efficiency. AMF-Net incorporates cross-attention to fuse global and local features, enabling the recursive generation of complete point clouds across multiple resolutions. Leveraging High-Performance Computing (HPC) environments, the model efficiently handles large-scale datasets, making it well-suited for parallel and distributed systems. We evaluate AMF-Net on the widely used PCN dataset, where it demonstrates superior performance in retaining structural fidelity and generating high-quality point cloud completions compared to existing methods. Through extensive ablation studies, we show that the multi-stage feature fusion module and attention mechanisms significantly contributes to the accuracy and detail of the completion results.