search

sniklaus / 3d-ken-burns

an implementation of 3D Ken Burns Effect from a Single Image using PyTorch
Other
1.52k stars 225 forks source link

How to Use the Refinement Network #44

Closed dfrumkin closed 4 years ago

dfrumkin commented 4 years ago

Hello Simon! If I want to run your refinement network on a given disparity map that was not obtained from your first network, what are the requirements on the input? Suppose the disparity map is normalized to 0..1, where 1 corresponds to the nearest object and 0 to the farthest one (alternatively 0..255), should I scale it in some way before feeding to the refinement network?

sniklaus commented 4 years ago

You can call it with tenDisparity = disparity_refinement(tenImage, tenDisparity). The guiding image needs to be four times the resolution of the disparity map. No need to shift or scale the disparity, the refinement network is invariant to shifts and scales. Let me know in case you are running into any issues.

dfrumkin commented 4 years ago

Thank you, Simon! I've played with it a little bit and indeed the difference after scale/shift is negligible. The only catch is that the values have to be non-negative.