MSP-Net: Multi-Scale Point Cloud Classification Network

The point cloud classification network is difficult to bring into full play the advantages of convolutional neural network. Aiming at the problem, a multi-scale point cloud classification network MSP-Net is proposed. Firstly, based on the requirements of completeness, adaptability, overlap and multi-scale characteristics of local area partition, a multi-scale local area partition algorithm is proposed. Based on the algorithm, multi-scale local areas are obtained by input of the point cloud model and its features of different levels. Then a multi-scale point cloud classification network including the modules of single-scale feature extraction, low-level feature aggregation and multi-scale feature fusion is constructed. The network fully simulates the principle of convolutional neural network, and has the basic characteristic that with the increase of network scale and depth, the local receptive field becomes larger and larger, and the feature abstraction becomes higher and higher. Experiments on the benchmarks ModelNet10 and ModelNet40 demonstrate the feasibility and validity of the proposed algorithm. The classification accuracy for above two datasets are 94.17% and 91.73% respectively, and the classification performance is on par or better than state-of-the-art methods.