Enhancing U-Net with low-rank attention skip block for 3D point cloud segmentation

The U-Net framework has been widely applied in the field of 3D point cloud segmentation. Most existing methods primarily focus on employing more powerful encoders to enhance the understanding of point cloud data. However, some inherent operations in traditional encoders, such as down-sampling and local neighborhood feature aggregation, restrict the effective receptive field of the model to a relatively small region. To address this issue, we introduce a global Attention Skip Block (ASBlock) to improve the skip connection in the U-Net framework, facilitating feature fusion between each point and the global context. Moreover, inspired by the dense-connection idea exploited by some well-known works such as DenseNet and 3D DenseNet, we further extend the proposed ASBlock to a new version with dense connection that can integrate more global contextual information for better segmentation performance. Additionally, to reduce computational complexity, we compress the global attention model using approximate low-rank matrix decomposition, leading to the development of the Low-rank Attention Skip Block (LrASBlock). This module can be efficiently applied to large-scale datasets and seamlessly integrated into existing U-Net segmentation networks as a universal plug-and-play tool. Finally, extensive experimental results on multiple datasets demonstrate that the integration of LrASBlock can significantly improve segmentation performance of several typical U-Net-based methods. Code is available at https://github.com/Ysc156/LrASBlock.