search

google / stereo-magnification

Code accompanying the SIGGRAPH 2018 paper "Stereo Magnification: Learning View Synthesis using Multiplane Images"
https://people.eecs.berkeley.edu/~tinghuiz/projects/mpi/
Apache License 2.0
390 stars 87 forks source link

Shifting image2 to image1 #28

Closed phongnhhn92 closed 4 years ago

phongnhhn92 commented 4 years ago

Hello guys, I just start reading the code of this repo and I can see that the first processing step is to shift source image(img2) to reference image(image1) by xshift and yshift. My question is: why we have to do that ? I have tried to find the reason on the paper but I found no explanation.

Also, do you shift img2 during training or is this just for the testing/evaluation ?

reyet commented 4 years ago

It's only necessary if you want to run on input images where the principal point is different between the two inputs. This is the case with some lightfield camera data, for example. We don't use it with the RealEstate10K data or during training.

phongnhhn92 commented 4 years ago

Thanks @reyet ! I will close this now.