MF 2 Net: A Multi-Level Feature Fusion Network for Point Cloud Segmentation

Segmentation of 3D point clouds finds extensive applications across various fields such as autonomous driving, robotics, and virtual reality. Due to the sparsity and irregularity of point clouds, the methods of normalizing point cloud arrangement may lead to the loss of spatial structural information. Point-based methods send point cloud directly to the neural network and extract features from the point cloud data by constructing an encoder-decoder network. However, these methods use the last layer of the decoded information as the output and do not capture the underlying semantic information. To address the above issues, a multi-level feature fusion network (MF2Net) is proposed for point cloud segmentation. At first, Point Cloud Self-Attention Module (PSAM) is proposed to utilize global information to weigh the importance of each point, which addresses the issue of independence between different points and help the network effectively exploits the representative semantic information. Then feature aggregation module (FAM) fuses these representative semantic information so that the different information between different layers can complement each other. After that, multi-output loss function is constructed to compensate for the loss of spatial characteristics in the encoding process. The experimental results demonstrate that the MF2Net exhibits excellent performance. The proposed method obtains a mIoU of 86.2% on the ShapeNet dataset and get the mIoU of 59.1% and 63.7% on large indoor and outdoor datasets.