Developer Center

Intro:
I've been working on this stitching project of mine for over a year now. It utilizes motion detectors to get the rotation matrixes representations, and then I convert them to camera space rotations by utilizing Eigen. The calculations work fine.
Problem:
The problem is that there's always noise in the motion detectors which can't be removed, which then reflects on the result - as shown in the image below.
Question:
Is there any way to minimize the errors shown in the image above by taking advantage of the parameters below? Or does anyone have any idea of how to "hide" the errors? Make them less noticeable?
Parameters:
Each image will be taken sequentially in a LTR order.
Each image will overlap the other image by 20%.
Each image will be separated by 20 degrees.
What I've tried:
SeamFinder - works ~OK for minimalistic errors, but, the error shown on the black monitor would turn into clippings and overall look weird.
Feature matching & extraction - I've tried to use feature matching and extractions etc, which often produces a better result than the image above, however, the downside is that it's very expensive to use.
What I think is a good idea, (feel free to argue against?):
Instead of SeamFinder, maybe utilize like a alpha gradient blending that blends each image by it's alpha value - thus making it less noticeable? Take a look at the image below.