Point clouds analysis with extreme learning machine and high order point CNN

Abstract

The thesis presents two novel algorithms for point clouds analysis. One algorithm is Point Clouds-based Extreme Learning Machine Auto-Encoder (PC-ELM-AE) which is the first network to incorporate Extreme Learning Machine (ELM) theory for processing point clouds. PC-ELM-AE stacks multiple local structure modules to inspect local structures and uses three ELM Auto-Encoders (ELM-AEs) to extract per-point features in hierarchy. This algorithm tremendously reduces time cost compared to deep learning methods and achieves decent performance on ModelNet40 and ModelNet10. The other algorithm is High Order Point CNN (HOPCNN) that is designed to generalize convolutions to unordered point clouds. HOPCNN combines bilinear pooling and trilinear pooling with convolutions to make convolutional outputs permutation-invariant and constrains ranks of convolutional kernels for efficient computation. HOPCNN can simultaneously benefit from high order statistics captured by bi/trilinear poolings and expressive capabilities of convolution operators. The algorithm achieves comparable or better performance than state-of-the-art on ModelNet40 and ShapeNetPart.