PCAPN: An enhanced feature extraction framework for point cloud

Point cloud is a widely used geometric data structure in the missions of 3D reconstruction, digital city and geologic survey etc. Extracting sufficient information from point key to deal with aforementioned missions. However, huge number points lead computational complexity inefficiency during training process. To this problem, paper proposes novel framework name Principal Component Analysis Net (PCAPN) for feature extraction cloud. Firstly, sampling module namely Sampling (CPS) designed generating several candidate sets by different scales, which defines centroids local regions. Secondly, based on MLP adopted extracting vectors set generated module. Finally, extracted are concatenated together, then put into fully connected layer classification. The proposed was evaluated 2 benchmarks, i.e. ShapeNet part dataset ModelNet40. experimental results show that our efficient robust. In particular, significantly better than those obtained state-of-the-art frameworks. Our network 4.6% more accurate PointNet 1.1% higher PointNet++.