Open erhuodaosi opened 4 years ago
If I'm not mistaken, this solution does not make differences between 'foreground' and 'background' while extracting features from the training sample (style image), it should substitute the features it's able to 'find' in the target image with those from the map of cross-correlated pieces of style image (training sample) -- it's unlikely that there's an option for non-linear transformations of picture data already implemented in this repo. Even if it was, at least two issues would have to be solved -- how do you separate deformation mask boundary from the 'style' features (color keying and similar techniques?) and what's your strategy for filling in the empty space outside of the deformation mask area after you perform the deformation?
There's not much I've found online regarding non-linear 2D transformations, though you might be interested by this:
https://people.engr.tamu.edu/schaefer/research/mls.pdf https://github.com/cxcxcxcx/imgwarp-opencv
@rabdulatipoff Hi,
Thank you for your advice!Thanks a lot! I believe it would help me a lot!
Thank you for your patient again.
Best wishes, FengLang
I'm sorry to bother you,but I have a question on could we apply fast-style-transfer to image deformation.
In style-transfer,we usually provide a content image and a style image to generate results.And now,could we give a content picture,and let it transform into a warpping image?If the answer is yes,what is the style image?How could we get the style image?
Here are two pictures:one is the original picture,and the other is the transformed image.
![after](https://user-images.githubusercontent.com/45588148/87241425-845a5c00-c455-11ea-86e1-f7b467042e9c.png)
I would very appreciate if you could help me,thank you!