Intrinsics

[bobluoluo]

The training data might could have different Ks, which is dependent on capture devices, cameras on cell phones. It seems to me that mean multiple K matrices( maybe even distortion coeffs) would be hard-coded to the network in the process. How does it affect the result? What are your thoughts about decoupling the intrinsics from this en/decoding network? Can intrinsics be provided when inference?

[yasaminjafarian]

We use a fixed K matrix. See here If your task should be aware of the camera parameters, I think decoupling should help.