SGReg: Swift and Robust Graph Matching based Point Cloud Registration
Wenhe Liu, Xingpu Wei, Haodong Wei, Lei Chen, Hao Zhang · 2024
Existing point cloud registration methods are computationally intensive in complex scenarios, making it difficult to achieve an ideal balance between accuracy and efficiency. To address this issue, we propose SGReg, a lightweight point cloud registration network. The core innovations of SGReg include the redesigned feature extraction network ReDGCNN, which employs multi-branch convolution and re-parameterization techniques; the introduction of a Simple Transformer (ST) as an edge generator, using a Linear SelfAttention mechanism (LSA) instead of traditional multi-head attention; and the innovative parallelization of the attention mechanism and MLP layers in the transformer structure. These optimizations significantly reduce computational complexity and inference speed while maintaining strong feature expression capabilities. Experiments show that SGReg outperforms existing methods in various benchmark tests.