swapping-autoencoder-pytorch
swapping-autoencoder-pytorch copied to clipboard
taesungp
how to keep the structure and extract feature?
Your work is so amazing. When I use the interpolation in the texture code, it will change the structure of source picture slightly. In the mountain dataset, for example, we can see the obvious change of mountain`s structure. But in this amazing picture:
The structure of tree and land can be remained unchanged. I try to add a freqency domain loss to constrain the structure, but the results are still not so good. So may I ask you how to keep the structure well and how to extract the green summer feature, winter snow feature or some other features just like in the above picture?
thank you for your reply!
Your work is so amazing. When I use the interpolation in the texture code, it will change the structure of source picture slightly. In the mountain dataset, for example, we can see the obvious change of mountain`s structure. But in this amazing picture:
The structure of tree and land can be remained unchanged. I try to add a freqency domain loss to constrain the structure, but the results are still not so good. So may I ask you how to keep the structure well and how to extract the green summer feature, winter snow feature or some other features just like in the above picture? thank you for your reply!
I think the training alogirhtm not actually constraint the structure code strictly, model is not guarantee the structure of translated image can be retained
See figure 6 in this paper, they discuss the effect of patch size selection, if use large patch size, the content will change significantly, I think using smaller patch size can preserve more structure (1/16~1/8)
That is correct, in that using smaller patch size will help. You can also make the structure code larger by reducing the number of downsampling steps in the encoder. (see this issue)
That is correct, in that using smaller patch size will help. You can also make the structure code larger by reducing the number of downsampling steps in the encoder. (see this issue)
How about the number of crops? If I use smaller crop size such as 1/16~1/8, do I need to change the number of crops from 8 to 16?
That is correct, in that using smaller patch size will help. You can also make the structure code larger by reducing the number of downsampling steps in the encoder. (see this issue)
How about the number of crops? If I use smaller crop size such as 1/16~1/8, do I need to change the number of crops from 8 to 16?