Perspective Transform of an Image

Hey
I don’t normally post questions as I like to solve things myself, but I have been stuck on this for over half a year. Its very much so an algorithm question, but is independent of OpenGL for the most part (other then I use OpenGL in my app) so please ignore this if it was inappropriate to post). I do a lot of basic image editing and decided most of the programs I used did far too much and cost a lot. That got me interested in building my own image application. I have built my own application and it does everything I need except for doing a proper perspective transform. I have tried some many different algorithms, I have researched the topic myself (and I am not the best at math, but I don’t think I am entirely stupid).

The issue I have is I want to be able to change the coordinates of one corner of the image (like I can in Gimp) and have none of the other image corners be affected.
Just an example, what I would like to be able to do is shown here:
http://onebulb.com/D2.png

In this image I have moved the bottom left corner of my image up and to the right. The other 3 corners of the image have remained in the same spot.

When I implement any version of this algorithm myself, I end up with:
http://onebulb.com/D1.png

In the above image I have only tried to move, the top left corner down and too the right, yet the bottom right has also been moved. I use a mac so I have also tried using the Core Image Unit Perspective transform, and the exact same thing happens. I move one corner and at least one other corner also moves. I have even tried duplicating the exact same matrix that Gimp or applications say they are using and I still have multiple corners that are moving.

Am I missing something here. Do I want more then a perspective transform. The algorithm I am currently trying to use is straight from Gimp 2.6. I do not see it doing anything fancy in particular but I am sure there is a lot happening behind the scenes that I am unaware of. On a Mac I have even tried taking the transform that gimp outputs putting to a Core Animation Layer(Mac Specific – though I have tried the exact same thing in OpenGL) and the result is still always the name, no matter which algorithm I use for my perspective transform and no matter how I apply it.

I have copied and pasted the gimp version here (mine is currently littered with commented out code as I experiment)
http://onebulb.com/gimpsrc.txt

So am I missing something? Am I looking at the wrong code in Gimp perhaps. Is there something extra I need to do to lock corners down? Anyone have an implementation of what I want that they are willing to share. Perhaps this is not entirely a perspective transform that I want?

Thanks for any help.

Here is rather old, but very good article about computing quadrilateral-to-rectangle perspective transformation:

http://www.comp.nus.edu.sg/~lowkl/publications/lowk_jgt2003.pdf

I must miss something, but I would not play with view frustum, I would just draw picture as a texture (since you use opengl) on a quad, moving vertices when the user is moving corresponding corners in the original picture. Then Opengl, will do all the job to obtain the 1st screenshot.

dletozeun,

You’re right, but if you made such a transform on arbitrary quadrilateral, you would see artifacts
http://www.r3.nu/~cass/qcoord/

Hey

Thanks, I will give that paper a read as well as the link. I think I started many months ago doing an arbitrary transform and saw artifacts and was quite puzzled. I was more interested in the math now though to know what im doing wrong so you have given me two new avenues to investigate. Thanks, though I am still puzzled as to what gimp does.

Thanks

Thanks Jackis, I did not realized it :slight_smile:

De rien, dletozeun :slight_smile: