MLFNet-Point Cloud Semantic Segmentation Convolution Network Based on Multi-scale Feature Fusion

In the semantic segmentation of a point cloud, if the spatial structure correlation between the input features and coordinates are not fully considered, a semantic segmentation error can occur. We propose a method of spatial convolution that makes full use of the characteristics of a multiscale spatial structure by combining local and global features. We call this method MLFNet. We also propose a multiscale feature framework. First, the point cloud is simplified by obtaining the weighted farthest point (by down-sampling combined with farthest-point sampling and the weighted average). The near-near domain of each sampling point is then obtained by a KK octant search (an octant search optimized by the k-nearest neighbor and a custom threshold), and feature information is obtained. The feature information is added to the subsequent multilayer perceptron, and fusion of local context information is achieved. Finally, the fusion features in multiple directions are maximally pooled. Our method was tested on self-made datasets and other standard basic datasets (ModelNet40, ShapeNet, and Stanford large-scale 3D indoor spaces (S3DIS) data). The accuracy of segmentation was 0.937 in our dataset; two percentage points higher than the latest deep learning method. Also, our method obtained a mean intersection over a union of 0.867 in ShapeNet, which was 0.3 percentage points higher than the latest PointGrid. The accuracy on S3DIS was 0.8153, which was three percentage points higher than the latest spatial aggregation net. The results of semantic segmentation verified the superiority of the proposed method.