FreezeAdaCRAFT Point Cloud Network

D point cloud completion aims to reconstruct complete object shapes from partial observations, addressing challenges such as data sparsity, missing structures, and the restoration of fine details. In this paper, we propose the FreezeAdaCRAFT Point Cloud Network (FAC-PCN), a novel framework that integrates adaptive frozen sampling with cross-scale Transformer-based feature enhancement. Specifically, the Adaptive Freeze-based Hybrid Sampling (AFSM) module selects key geometric points to improve feature stability, while the Multi-Scale Cross-Resolution Feature Augmented Transformer (MS-CRAFT) module enhances semantic interactions across different scales. A cascaded generation module then progressively upsamples the seed points to produce complete and high-fidelity point clouds. Experiments conducted on the PCN dataset demonstrate that FAC-PCN significantly outperforms existing methods, particularly in reconstructing complex structures and preserving fine-grained details.