Abstract
Geometric camera calibration is often required for applications thatunderstand the perspective of the image. We propose perspective fields as arepresentation that models the local perspective properties of an image.Perspective Fields contain per-pixel information about the camera view,parameterized as an up vector and a latitude value. This representation has anumber of advantages as it makes minimal assumptions about the camera model andis invariant or equivariant to common image editing operations like cropping,warping, and rotation. It is also more interpretable and aligned with humanperception. We train a neural network to predict Perspective Fields and thepredicted Perspective Fields can be converted to calibration parameters easily.We demonstrate the robustness of our approach under various scenarios comparedwith camera calibration-based methods and show example applications in imagecompositing.