I first separate pictures into defined DOF part and the rest. I start blurring with the part of picture without DOF (replaced with black background), and iteratively blur it, so the part of the picture that is further from focus line have more blur. Then, I add back DOF to blurred picture to generate the final result. Complex DOF regions are defined in each picture using MATLAB roipoly and ginput function. But, the edge of each mask will not fit to the rest of the picture naturally, so I combined it with the standard horizonta/vertical DOF to make it appear to be natural.
Color saturation channel of pictures are tuned for the better visual effects.
In general, tilt effect works better on if object within DOF is small. Usually the picture of streets and crowd can be a good picture to represent with tilt shift. The pictures have "depth" also looks better with tilt shift. Compared to the complex DOF, simple horizontal or vertical DOF appear to be more natural on some pictures.