A multi-scale point cloud partitioning method by geodesic distance-based
Point cloud objects are typically composed of complex,unordered,non-Euclidean spatial structures and multi-scale features,often being dynamic and unpredictable.Current self-attention modules rely on dot products and matrix transformations,which make adequately capturing the multi-scale non-Euclidean structure of point cloud objects difficult.This paper proposes Geodesic-based Multi-scale point cloud partitioning Transformer(GMT),a new multi-scale self-attention module to solve the above issue,aiming to extract non-Euclidean geometric information from point clouds at multiple scales to enhance their representation capabilities.GMT achieved an overall accuracy of 93.2％and a mean accuracy of 90.5％on the ModelNet40 dataset.On the ScanObjectNN dataset,it achieved an overall accuracy of 82.5％and a mean accuracy of 81.1％.Compared to other mainstream methods such as Point-TnT,GMT yielded competitive results.Experimental results indicate that GMT possesses a strong capability to capture non-geometric features to some extent.
NETL
NSTL
万方数据
