Code: https://github.com/Yochengliu/Relation-Shape-CNN
文章：Relation-Shape Convolutional Neural Network for Point Cloud Analysis [arXiv] [CVF]
This paper is CVPR 2019 Oral & Best paper finalist
Abstract & Introduction
in the point cloud, due to the irregular shape of the implicit point difficult to capture, such that the cloud point of the analysis is very challenging. In this paper, the authors point cloud data for the proposed RS-CNN, namely Relation-Shape Convolutional Neural Network, the core idea is to learn from the point cloud geometry topology relationship information. RS-CNN in multiple data sets have made SOTA performance.
The main contributions are as follows:
A novel study from the convolution operator relations relationship shape convolution. It can be explicitly point geometry of encoding, which largely improve the perception of shape and robustness. Is simply, RS-Conv
made a deep hierarchy has a relationship convolution shape, that is, RS-CNN. CNN will be extended to a regular grid irregular configuration, point cloud context shapes perceptual learning. Is simply, RS-CNN-based RS-Conv designed
in three task of challenging benchmark extensive experimental and theoretical analysis and in-depth experience, proven RS-CNN reached the level of SOTA. Is simply good effect high precision (modelnet40 93.6)
the Shape Representation-Aware Learning
First, induction of a general convolution formula
\ [f_ (P_ (sub) ) = \ sigma (A ({T (f_x_j), \ forall x_j \ in N (x_i
)})), N (x_i) = {x_j \ mid d (x_i, x_j) <r} \] in 2D (i.e., image),
It represents the summation for the activation function
which is
