SDA-Net: A global feature point cloud completion network based on serialization and dual attention

Point cloud completion is essential for restoring 3D geometric data lost due to occlusions or sensor limitations. Existing methods often rely on k-nearest neighbor(KNN)-based local feature extraction, which focuses on neighborhoods around central points while neglecting critical global structural information. Additionally, Transformer-based approaches, while effective at modeling global context, typically use central point feature sequences to reduce computational complexity. This windowed attention strategy compromises the preservation of global context, leading to incomplete modeling of the point cloud's overall structure. To address these challenges, we propose SDA-Net, a dual-attention point cloud completion network utilizing multiple serialization strategies. These strategies transform unordered point clouds into structured sequences, enabling comprehensive modeling of inter-point relationships. Additionally, the dual-attention mechanism enhances global feature extraction through complementary spatial and channel-wise self-attention, effectively compensating for the loss of global context. Extensive experiments demonstrate that SDA-Net achieves state-of-the-art performance, including an average Chamfer Distance (CD) of 6.48 on the PCN dataset. Furthermore, it excels in real-world applications, accurately reconstructing fine-grained details in LiDAR-scanned point clouds. The source code is available at https://github.com/Hibiki-Ula/SDA-Net.