Closed minshu-kim closed 2 years ago
In fact, I'm not sure what you are confused with.
We estimate the homo on the resolution of 128X128, then this homo is scaled to a new homo corresponding to the original resolution. The reconstruction module also adopts the original resolution.
Excuse me again, My question is about the scale factor in the low-resolution branch in reconstruction module. In the reconstruction module, the image is always downsampled in 1/2 scale at Low-resolution branch?
Actually, the warped images are first downsampled to 256*256 (not 1/2 scale, because the resolution of the warped images is arbitrary) first. In this low-resolution, the misalignments can be reconstructed better. Then, we use this low-resolution reconstructed result to facilitate the reconstruction in high-resolution.
I understand. Thanks for your clear explanation :)
When the output resolution of alignment network is 128X128 or 256X256, what is the resolution do you use to downsample the output of alignment network?