How can you transform any image from the left one to the right?
I want to do this WITHOUT choosing 4 reference points, but i DO have access to metadata like roll, pitch and yaw
>>8080466
if it is modelled in 3d and you have roll/pitch/yaw you just set 2 of those to 0 and one to 90
What happens if you just pick the edges of the first image and apply the same transformation?
>>8080488
OP here:
I need to do this for multiple images, which all have different roll/pitch/yaw.
I want to create a birds eye view, so I don't know if what you're suggesting would help
I suggest gimp
try quaternions
Do you have the field of view?
>>8080558
Nope, just roll/pitch/yaw and altitude (I'm working with images from a UAV)
Is there a reason to not do it with reference points? Have you ever used MS' picture processing? I think it does exactly that.
>>8080726
I've got 400 images for which I want the transformation matrix.
I thought it would be easy to calculate this, given the roll/pitch/yaw but it's giving me a headache
This is image rectification, go read what Hartley said
So you need a 3d roto matrix, look at euler angles
It could also bo modeled with a perspective transformation, its 7dof 3*3 matrix
>>8080466
Unless it's a 3D model, there is no way to do this easily without a specialist program
OP the word you're looking for is homography.
>>8080466
use roll and pitch as your theta and phi and set ro to 1 and you have a unit normal vector. assume your plane contains the origin and just choose arbitrary reference points on it.
>>8080466
I don't actually think that's possible. Consider the case where the left image is not actually a plane which is flat relative to the ground, but is actually a trapezoid directly facing the view angle. These two cases would look very different from directly above!
Without some sense of the distance of various points from the camera, perspective transformations are impossible. You could certainly *assume* everything was just 2D images painted on a perfectly flat ground plane and warp them accordingly, but unless you make this assumption it cannot be done without more data than a single image and pitch/roll/yaw.