image background subraction

I'm tired to argue over this topic with people who saw many practical examples of such work but do not understand the essence of the problem. Solving this problem with 100% fidelity is theoretically impossible. This is one of the ill-posed problem. There are several reasons for that, but one of the most simple reason is this: there are no reliable criteria to tell a background from a foreground, there is no way to strictly define it. There is a number of other difficulties, some of them are fundamental, some are practical. Strictly speaking, to render modified scene, you need a complete 3D model of the scene and lighting, and on a photograph, a lot of information is already missing.

Now, as some practical solution do exist, the question is: are they good enough? Here is my answer: it depends on who is looking. If a person is experienced in photo editing or even in painting, she or he will detect an forgery in no time. Take myself: I'm not very experienced in photo editing, but in near past I spent a good time at photo editors. And what? I can detect false transformation in all high-quality images performed by any automatic methods almost immediately. At the same time, I can fail to see the forgery when it is done manually by a pretty good master. We human being are quite good in image recognition based on usual assumptions done on the image, so best of us can do the forgery very well. Not the automatic processing. And real masters can detect the forgery even if they are done by a photo forgery master.

At the same time, on low-quality images the forgery is hard to spot. If you want sloppy cheesy low-resolution or low-quality images, you can try to go in for it (I actually tried, and it was not to worst.) You see, the problem is too difficult to hope for a quick and dirty solution just our of curiosity — it will still require some hard work. From the other hands, the expected results would be too boring to bother about them. My advice would be: don't bother!

—SA