Abstract
TL;DR: This image stitching algorithm utilizes feature matches between images to compute an homography matrix. Using this homography the second image is warped, and then the two images are layed on top of one another. Then, dynamic weighted fusion is used to remove the seam line caused by the pasting of the images onto one another. The result is a stitched image with little to no seam line present.
Above is the two original images taken side by side. And below is the stitched result. As you can see there is little to no seam line present from the image and the alignment is as it should be.
Pipeline overview
The pipeline of the image stitching consists of main parts.
The first is converting the images to grayscale so that SIFT can extract matches.
After extracting matches OpenCV's FlannBasedMatcher is used and then Lowe's ratio test is performed
to ensure good matches. These matches are displayed in the iamge below
Based on the extracted matches an homography is retrieved and this is applied to the right image to
warp it in line with the left image.
The left image is then added to the warped right image. This will create a very sharp
vertical seam line at the right hand edge of the left image. Although in this scenario the
seam line is very minimal, it is very evident in the areas pointed out by the arrows.
After applying the weighted fusion on the image the seam line in removed and the result is a
blending of the left image and right image.
The final step is to fix the corners as a result of the image warped. Applying a simple warp of the image
and moving the corners to the edges fixes this and gives the final image.
