Point Clouds Classification and Segmentation Based on Local Fusion Network

The irregularity and sparsity of point clouds make it challenging to directly process raw 3D points using existing deep learning frameworks for classification and segmentation tasks. Aiming to solve this challenge, an end-to-end learning network is proposed. The key component of the proposed network is Local Fusion Block (LFB), which can be easily embedded into other networks. The function of the block is to fully integrate features from local points and a single point. By stacking the blocks, the network can obtain a larger receptive field, and achieve an effect similar to image convolution. Experiments on two representative shape classification and part segmentation datasets show that the proposed method achieves competitive results compared with some representative approaches in terms of accuracy.