Residual Blocks PointNet: A novel faster PointNet framework for segmentation and estimated pose

摘要

Given recent advances in Segmentation of Convolutional Neural Networks (CNNs), this paper aims to propose a more efficient structure which directly consumes point clouds for segmentation and estimated pose. More specifically, a novel Residual Blocks PointNet is proposed providing a fast framework taking point sets as input and predicting 3D object part segmentation and 3D pose. The network of the proposed structure has been established composed of two subnet works: a key branch for 3D object part segmentation and the other branch for spatial transform to predict a 3D affine matrix. The major branch contains more residual blocks, which encapsulate shortcut connects with specified layer numbers, growth rate and conv (1*1)-bn-relu structure. The key point is the decrease of each level of network computing and the reuse of feature maps. The other is a parallel classification network for estimated pose with share portion weight except 3 groups of full connected layers. Empirically, Residual Blocks PointNet shows faster rate of convergence and acceptable performance.