CVPR文章

• Occlusion Aware Unsupervised Learning of Optical Flow

@misc{wang2018occlusion,
      title={Occlusion Aware Unsupervised Learning of Optical Flow}, 
      author={Yang Wang and Yi Yang and Zhenheng Yang and Liang Zhao and Peng Wang and Wei Xu},
      year={2018},
      eprint={1711.05890},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

分别求出前向光流和后向光流，通过前向光流进行重建后向光流 $\widetilde{I}_1$ 向后流用于通过前向翘曲产生遮挡贴图$（O）$ photometric loss 光度损失：相当于求两个图片中所有点的相似性。 smoothness loss 正则化平滑：仅基于光度损失的无监督学习对于无纹理的地方是模糊的。减少模糊度最常用的方法就是平滑正则化smoothness loss函数。

文章提高了一种端到端的无监督学习框架，可以训练为标记视频的光流信息。主要参考了FlowNets模型，做出了一定的改进。

可用思路:将FlowNets模型用于我们实验室的PWC-Net模型。其他地方基本可以不用做改变。可以看看后面其他论文改进的部分。

• Learning by Analogy: Reliable Supervision from Transformations for Unsupervised Optical Flow Estimation

@inproceedings{liu2020learning,
   title = {Learning by Analogy: Reliable Supervision from Transformations for Unsupervised Optical Flow Estimation},
   author = {Liu, Liang and Zhang, Jiangning and He, Ruifei and Liu, Yong and Wang, Yabiao and Tai, Ying and Luo, Donghao and Wang, Chengjie and Li, Jilin and Huang, Feiyue},
   booktitle = {IEEE Conference on Computer Vision and Pattern Recognition(CVPR)},
   year = {2020}
}

和上一个网络有一定的差别 以PWC-Net 网络图： 成功在自己笔记本上实现部署模型并且测试代码： 模型部署比较麻烦。需要在cuda9.0上运行。 思路同样可以借鉴。并且可以直接将超声B模式的图像整理以后直接拿去训练。

** What Matters in Unsupervised Optical Flow**

@article{DBLP:journals/corr/abs-2006-04902,
  author    = {Rico Jonschkowski and
               Austin Stone and
               Jonathan T. Barron and
               Ariel Gordon and
               Kurt Konolige and
               Anelia Angelova},
  title     = {What Matters in Unsupervised Optical Flow},
  journal   = {CoRR},
  volume    = {abs/2006.04902},
  year      = {2020},
  url       = {https://arxiv.org/abs/2006.04902},
  archivePrefix = {arXiv},
  eprint    = {2006.04902},
  timestamp = {Fri, 12 Jun 2020 14:02:57 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2006-04902.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

这篇论文主要对Unsupervised Optical Flow涉及到的一些常见模块进行实验分析，有很好的指导意义。