sniper's profile - activity

I have get projection matrix from homography and camera parameter for AR. For checking if the result is good, I tested if ROI (0, 0) (w, 0), (w, h), (0,h) in the coordinate of reference image can be visualized in different image at different view. I assumed that ROI 4 points have zero depth. The result is successful.

Thesedays, I implemented the same function from fundamental matrix. 1. get Fundamental Matrix 2. Essential matrix from Fundamental 3. decompose essential mat into rotation & translation matrix 4. get projection matrix from camera parameter & rotation & translation Matrix.

Like the case I used projection mat from homography, I transformed 4 ROI in reference into the other image at different view. But the result is different from one by homography. Non feasible result was shown as the strange ROI box.

I was wonderingn if the projection matrix by homography is different from one by fundamental matrix?

please let me know how to transform ROI in reference image into the same scenen at different view.

For explanation, assume that we have two image (image 1 and image 2) and camera intrinsic matrix K. We can get the correspondence between two image by local feature matching. What I want to get is 6DOF pose from image 1 to image 2.

To get pose, I got fundamental matrix F from 2D-to-2D correspondence. Then, I calculated essential matrix E from F and K. After decomposing E, finally I made projection Matrix P. I would like to transform Rectangle ROI in image 1 to ROI in image 2. I set depth of ROI in image 1 to zero. (0, 0, 0), (w, 0, 0), (w, h, 0), (0, h, 0) Then, to get ROI in image2, I used project matrix P. But ROI is strange. Is there anyone to give some guideline?

I read the concept of POSIT. But I think that it is not suitable. I want to estimate of non planar object, but 3D model (or 3D points array) is not given. What I can get are single image about non planar object & correspondences. Wikipedia said that POSIT need 3D points array model and their correspondence in 2D image. It is not same to my condition.

Thank you very much. I wonder if POSIT can also estimate pose of planar object? Now I do not use intrinsic camera matrix since it need not real 6 dof pose but transformed roi. POSIT need intrinsic camera matrix? If so, we can get approximated intrinsic camera matrix given image resolution?

I am implementing 6DOF pose estimation of object. First, my implementation run successfully in the case that given target template object is planar structure. But, my algorithm can not track object which is non planar structure since it use homography for pose estimation (Actually, I need not 6 DOF pose but transformed rect ROI from original template ROI when camera is moved)

When we assume that one image with non-planar structure object is given as template image, Is there method that can get transformed ROI of the template when camera is moved?