E-Mamba: An efficient Mamba point cloud analysis method with enhanced feature representation

Abstract

As a key technology for three-dimensional space analysis, point cloud analysis is widely used in many fields such as automated machinery, unmanned vehicles and virtual reality. Learning local and global features of point cloud is crucial for gaining a deep understanding of point cloud data. In point cloud local feature learning, sub-clouds with center coordinates subtracted are usually used as point patches, which are then input into mini-PointNet to enhance the point cloud feature representation. However, this method has a high dependence on the point cloud density, which affects the model performance. In this work, we introduce E-Mamba, a new model for efficient point cloud analysis. We use Scalable Embedding to rescale and patch embedding sub-clouds, which improves the model’s feature representation and generalization capabilities for point cloud. In addition, we also introduced Holosync Reordering Pooling to reorder tokens while preserving the original sequence, and used the hybrid pooling method to extract global features. In this way, the model fully utilizes the periodicity of Mamba while achieving good generalization and global feature extraction capabilities. We conduct extensive experiments on ModelNet40, ScanObjectNN, and ShapeNetPart datasets. The results show that E-Mamba can achieve superior performance while significantly reducing GPU memory usage and FLOPs, whether pre-trained or not.