PointWave-MLP: Point Cloud Analysis Based on Wave-MLP Architecture

Due to the irregularity and disorder of point clouds, point cloud analysis based on deep learning remains a challenging task. Although the previous point cloud analysis networks based on multi-layer perceptions (MLPs) exhibit a simple structure, they fail to consider semantic information among different tokens by exploiting fully connected layers with fixed weights for token aggregation. To address this limitation, we draw inspiration from the Wave-MLP architecture in the 2D image field and introduce the concept of phase-aware token mixing into 3D point cloud processing, aiming to dynamically adjust the relationship between tokens and the fixed weights. In this paper, we propose a novel point cloud analysis network called PointWave-MLP. Experimental results demonstrate that PointWave-MLP achieves significant improvements in point cloud classification and segmentation tasks, outperforming some CNN-based and Transformer-based methods.