An Enhanced Downsampling Transformer Network for Point Cloud Semantic Segmentation

In outdoor environments, point cloud collection is often affected by external factors, leading to noisy and outlier points. Despite various pre-processing methods, noise is difficult to completely remove from training data. In addition, due to the disorderly nature of point clouds, traditional models have difficulty in extracting local information from point clouds, resulting in low accuracy. Therefore, we propose an enhanced down-sampling Transformer network for point cloud processing. Firstly, to tackle the influence of noise on feature extraction, we propose an enhanced down-sampling method, which constructs a graph by connecting the central point and its neighbors after down-sampling, and trains the weight between points to eliminate the effect of noise on feature aggregation. Secondly, after down-sampling, we use a novel Transformer module to train and update point cloud features, which includes residual modules and MLP to adapt to our down-sampling module and prevent model overfitting. Finally, extensive experiments are conducted on the S3DIS dataset, and the Mean Intersection over Union reaches 64.5%, indicating that our model is highly competitive in point cloud processing.