out of frame landmark detection with CNN

a robust landmark detector must cope with occlusion (the landmark surrounding are occluded by another object) the detection can still be performed but it should be stressed that the landmark is occluded. below you can notice that the occluded landmarks are presented in red

in my opinion, there is a more challenging problem that I have not found a lot about in the literature which is the case where the landmark area is out of frame.
imagine that you face the case below where the right eye extremity is out of frame but you model will spit any way the same number of the detected landmark.

how can one manage this in the training step? how could we inform our model that the landmark is out of frame?