Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes

本文介绍一篇ECCV2020里面关于点云无监督学习的文章。
论文
目前还没有开源代码

1. 问题

点云形状的无监督学习

2. 思想

【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020
本文的主体思路是延续 Unsupervised cycle-consistent deformation for shape matching 这篇文章,在其基础上添加了sinkhorn regularization模块,从而尽可能的使cycle correspondences成为bijiective map(也叫一一映射)。背后的assumption应该是,如果网络能够将一个点集在不同transformation下的correspondece找到,那么也应该学习到了discriminative pointwise features。 这种assumption和CVPR 2020 GraphTER的思想不谋而合,个人觉得GraphTER对此assumptaion的阐述更详细,更有说服力。

3. 算法

3.1 Unsupervised loss based on cycle consistency

Correspondence query between point cloud shapes.

首先给出确定correspondence的准则
【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020
但是由于arg max不可微,采用a scaled softmax function代替
【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020
Cycle-consistency loss.
【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020
接下来定义D距离函数
【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020

3.2 Sinkhorn regularization for bijective constraint

通过sinkhorn regularization 来保证一一映射的对应关系

对于一个严格的一一映射,【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020矩阵应该是一个单位阵。但是这个限定太强了,需要将其松弛为双随机矩阵(行和列和为1)。
【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020
【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020

4 实验结果

与Unsupervised cycle-consistent deformation for shape matching 这篇文章不同,本文采用了不同的setting, 实验总体更倾向于key-point detection, registration。只是在最后进行了一个全监督的segmentation测试。
【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020【点云识别】Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes ECCV2020
我对table 2的实验有些小疑问,等等看作者是否会 开源代码。

总结

文章的语言不错,以19年的文章为主体思想,结合sinkhorn regularization 完成了自己的work。个人猜测应该是cvpr 改投的文章,老师在其中应该起了不小的作用。估计这个框架训练起来时间要蛮长的,期待开源代码。

相关文章:

  • 2021-04-17
  • 2021-10-11
  • 2021-05-08
  • 2021-04-03
  • 2021-12-20
  • 2021-06-01
  • 2021-12-02
  • 2021-11-27
猜你喜欢
  • 2021-03-26
  • 2021-08-31
  • 2021-08-24
  • 2021-08-06
  • 2021-09-05
  • 2022-01-07
  • 2022-01-14
相关资源
相似解决方案