MLCompletion: Point Cloud Completion via Multiview Feature Augmentation and Progressive Refinement

Abstract

Point cloud completion is a fundamental task in 3-D sensing and vision, yet it remains challenging when partial observations suffer from severe structural missing regions and ambiguous local geometry. To address these issues, we propose MLCompletion, a point-cloud-only completion framework that incorporates self-generated multiview geometric augmentation and progressive shape refinement without relying on external images. Specifically, the proposed geometry-aware fusion module (GFM) projects the partial point cloud into three orthogonal depth maps and fuses the extracted depth features with 3-D point features, thereby enhancing global shape perception and missing-region representation. Based on this representation, the shape completion module (SCP) progressively refines the coarse prediction in a coarse-to-fine manner. Within the SCP, the dynamic feature fusion gate (DFFG) adaptively balances local geometric details and global structural context to improve reconstruction fidelity. Experiments on ShapeNet-55/34, PCN, and KITTI demonstrate that MLCompletion achieves competitive completion performance and consistently improves chamfer distance (CD) over strong baselines. Additional ablation studies, efficiency analysis, the DFFG interpretability results, and extreme sparsity tests further validate the effectiveness of the proposed design while revealing its limitations under highly challenging incomplete inputs.