search

JirongZhang / DeepHomography

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

Question about the image transformation range #9

Open zhangliukun opened 4 years ago

zhangliukun commented 4 years ago

In your project's homepage, the differences between the source images and the target images are small, most are the translation. I wonder whether your method could deal with the large transform such as random affine transform.

steve13durant commented 3 years ago

In your project's homepage, the differences between the source images and the target images are small, most are the translation. I wonder whether your method could deal with the large transform such as random affine transform.

I have the same question

zhangliukun commented 3 years ago

In your project's homepage, the differences between the source images and the target images are small, most are the translation. I wonder whether your method could deal with the large transform such as random affine transform.

I have the same question

In my opinion, this network structure cann't deal with the large transform because the feature map $F_a$ and $F_b$ are concated and then fed into the homography estimater. If two image pairs $I_a1$,$I_b1$ and $I_a1$,$I_b1$ are different, though the ground truth homography matrix is same, the results estimated by the network may be different. As the author said in the paper, this work is not able to be applied to such problems like image stitching.