Improved DGCNN Based on Transformer for Point Cloud Segmentation

Semantic segmentation task is an important branch of computer vision field. We propose a new model based on DGCNN [3] and Transfromer [6]. DGCNN is an excellent model that has achieved good results in semantic segmentation tasks. However, there are still some shortcomings, so we propose two blocks feature reinforcement block (FRB) and transformer feature block (TFB). The global feature is very important to the model, but in DGCNN it is only obtained through symmetric functions. Therefore, we use both FRB and TFB based on Transformer to improve the original model, because the self-attention block in Transformer has a natural advantage for global feature extraction. In the FRB block, we use the local features extracted by the Edgeconv block and the address coding information to extract the global features. In the TFB block, we integrate the output features of the first three stages to enrich the semantic information of the features. The model was tested on the Dataset Stanford Large-Scale 3D Indoor Spaces Dataset (S3DIS) [5]. The three metrics improved, Mean Intersection over Union (mIoU), overall accuracy (OA) and mean accuracy (mAcc). The mIoU increased by 2%, OA increased by 3%, and mAcc increased by 1.3% compared to the original model.