rows (y) and columns (x) are with respect to the camera sensor readout. in contrast, horizontal and vertical are with respect to the 'real world' (or 'scene').

'Normally' the sensors are oriented such that x respresents 'horizontal' while y represents 'vertical'. But if you rotate the camera (sensor) by 90 degrees, this relation changes.

Hope this helps

If You do not know the orientation of the sensor You can try to find the orientation on the basis of some priors. For example you can try to find the horizon if it is in the scene

I think it is according to the Camera (vision sensor) orientation. Usually photos are rectangular. Now Horizontal photographs are photographs that are wider than they are tall. Vertical photographs are photographs that are taller than they are wide. The orientation of an object inside an image is then decided with reference to the orientation of the original image.

I hope this is what you mean.

I think need some "point" on image that need for become "horizontal" point or "vertical" point. the image need the example for the "right" direction of image, u need classified it if u want it to use on face,car, or other. but if scenery will hard but we can use the sun as the "point" of vertical(the problem is lamp)

Well it depends on how you perceive the terminations or end points of line segments in an image. Sometimes these textural features are clear, and one can draw a distinct perception on the directionality in the image. However, this could be an ideal case, and the real world problem is much more complex. In analogy to fuzzy logic, images could combine characteristics from both horizontal and vertical; and in some cases the orientations could be closer (i.e. more probable) to be considered horizontal rather than vertical, and vice versa.