search

JirongZhang / DeepHomography

Content-Aware Unsupervised Deep Homography Estimation
MIT License
339 stars 57 forks source link

Some questions about paper #4

Open tomjerrygithub opened 4 years ago

tomjerrygithub commented 4 years ago

Hi,Mr.Zhang ，i have some questions about the paper，can you explain it ？ thanks~ 1、image input size is 315 × 560 ？ 2、How to get Ga with Ma and Fa ? 3、And then ， np.dstack Ga and Gb to Resnet34 ？ 4、Training data is imageA imageB and Homography matrix from imageA to imageB ? 5、Can the training set be generated like 《Deep image homography estimation》?

I want to try to implement your paper, but I feel that there will be some difficulties in the loss function part, can I get your help?

JirongZhang commented 4 years ago
  1. yes
  2. pointwise multiply
  3. torch.cat , do not use numpy
  4. Hab = imageA to imageB
  5. no All processes need to be implemented with torch or tf, so that the entire calculation process is differentiable. You need to read ’ Unsupervised deep homography: A fast and robust homography estimation model ‘(https://arxiv.org/abs/1709.03966) first。