MATNet: Semantic segmentation of 3D point clouds with multiscale adaptive transformer

In recent years, the Transformer model has made significant progress in semantic segmentation tasks. However, existing self-attention mechanisms perform well in capturing remote dependencies and global features, but ignore local area information in point cloud data and have limitations in dealing with multi-scale features. To address this problem, this paper introduces a multiscale self-attention fusion (MSA) module, which adaptively fuses features within different scale neighborhoods and learns global contextual features by connecting local neighborhoods. Then, the multiscale channel aggregation module (MCA)is used to perform deep point-by-point and point-by-point convolution of the point cloud channel, aggregating channel features at multiple scales to extract more accurate local feature information. Finally, in this study, the multiscale adaptive fusion (MSA) module and the multiscale channel aggregation (MCA) module form a sequential network structure that adaptively and dynamically adjusts different scales of point cloud objects to enhance the perception of objects of different sizes for better segmentation performance. By testing and validating the model on the publicly available S3DIS Area 5 dataset and the ScantNetV2 dataset, the model achieves mIoU index values of 71.9% and 72.3%, respectively, which demonstrates the effectiveness and superiority of the proposed method. Code will be made publicly available at https://github.com/Cocoyufei/MAT/tree/master.