A Lightweight Convolutional Dual-Channel Transformer Framework for Efficient Part Segmentation of Point Cloud Data

In recent years, transformer-based models have made significant breakthroughs in natural language processing and computer vision. However, these models have encountered challenges when dealing with point cloud data because the irregular and disordered structure of point cloud data leads to a huge computational and memory burden. To address this problem, this paper proposes a Multidimensional Convolution-Dual Channel Transformer Network for efficient processing of point cloud data. The MCDTN framework consists of two branches: the main channel enhances the modeling of cross-channel features through dynamic attention to optimize feature representation; the auxiliary channel further improves fine-grained segmentation capabilities through encoders, multi-scale information interaction, and spatial attention. Experimental results show that MCDTN performs excellently in shape classification and part segmentation tasks, effectively reducing computational costs.