twhui/LiteFlowNet

Question about data augmentation?

PeiqinSun opened this issue · 2 comments

PeiqinSun commented

hi. (我猜你应该**人, 因为我看到了香港中文大学.) 首先, 非常感觉你做的工作. 我目前也在研究关于CNN opticalFlow, 我有一些insight想和你讨论一下.

  1. flowNet的数据增广好像特别难以模拟出来, 至少感觉目前这一分支一直沿用caffe框架的原因我猜应该也是这个增广问题吧? 至少, 我用python复现的网络, 都在sintel上达不到原来用caffe训练得到的结果.
  2. 网络的一些特殊层真的那么重要吗? 我怎么感觉从我个人的实验来看, 除了残差金字塔的结构有明显作用, correlation和warpLayer的作用感觉并不是特别明显.

希望能和你讨论一下关于数据增广方面的细节问题, 如果方便, 能否有邮箱交流?

twhui commented

I simply used the augmentation layer as FlowNet2. I have not tried other alteratives other than Caffe.

I have no idea why you have such a conclusion. All the proposed components are important for optical flow estimation as presented in the ablation study.

Crow77 commented

How do you create de training.list and the test.list from the downloaded dataset (e.g KITTI12) ?
I'm trying to run the make-lmdbs-train.sh script

Please help