LONet: Local Optimization Network for 3D point cloud semantic segmentation

In the point cloud segmentation tasks, the existing local feature aggregation methods in the up-sampling and down-sampling stages still rely on Euclidean distance to constrain the local aggregation process. However, this approach is susceptible to the influence of abnormal points leading to inaccuracies in fitting the original geometric shape of the point cloud. Therefore, this paper proposes a local gradient aggregation module, which incorporates gradient information of neighboring points during the aggregation process. This enables the model to capture fine-grained geometric information and extract richer local features. Additionally, we introduce a symmetric sampling strategy to improve computational efficiency. The same original mapping indices were used for both up-sampling and down-sampling aggregation. Thus, a large number of additional k-nearest neighbor calculations are avoided. Furthermore, this paper introduces a position-aware encoding in the attention mechanism to address the positional cues for short-term and long-term contexts, facilitating positional-aware communication between points. Numerous comparative experiments prove the effectiveness of the method in this letter. It obtained 72.1% mIoU on ScannetV2, and 72.4% mIoU on S3DIS.