For networks determining direction like head pose and gaze estimation, what's the best way to compensate for an offset mounted camera (not having the camera straight in front of the face)?
Calibration with a "chess board" gives a transformation, but is it better to try and warp the input image or to adjust the network output vector?