GLoM-PCNet: A Deep Network for Multi-scale Information Fusion for Point Cloud Completion

Point cloud completion is a pivotal task within point cloud analysis, with the objective of transforming incomplete and sparse point clouds into their dense and complete 3D shape. Conventional methods often falter in effectively harnessing and synthesizing multi-scale, global, and local information during the reconstruction process. To surmount these hurdles, we propose GLoM-PCNet, a transformer-based framework to efficiently extract and amalgamate global and local multi-scale information for point cloud completion. This framework processes the input point cloud by employing a trilayered self-attention mechanism across different scales, which fosters dynamic interactions with global information and ensures the proficient fusion of multi-scale data. Furthermore, we innovate a k-nearest neighbor knn-based local multi-scale perception technique that significantly enhances the representation of local detail features through multi-scale integration. The harmonious fusion of multi-scale features at both global and local levels empowers GLoM-PCNet to achieve exceptional performance across a spectrum of benchmark datasets.