CODEWITHSUNDEEP
pythonopencvimage-processingcomputer-vision
syfy
syfy

Reputation: 647

Stitched image not well aligned, hence leading to duplication

As shown in the image below, one image is not well aligned with another, which is causing duplication of vessels. How do I get rid of the duplication?

enter image description here

Th way I'm currently going about stitching is, I first find the keypoints using SIFT, I then use flannbasedmatcher to match the keypoints, find the homography, and then warp both; the stitched, and the to-bo-stitched image.

Upvotes: 0

Views: 323

Answers (1)

Micka
Micka

Reputation: 20130

homography isnt perfect for non-planar scene or non-pure-rotation-camera, so there will always be errors SOMEWHERE in the aligned images.

Imho best to do (if you have to use homographies) is to distribute the errors over the whole image by making sure that the points used to compute the homography are distributed well over the whole image.

Or use a blending method that hides the duplicated parts in either of the images...

EDIT:

Here is a paper describing how to distribute the keypoints

http://www.lfb.rwth-aachen.de/bibtexupload/pdf/BEH10g.pdf

They claim to reduce the error.

EDIT 2:

another approach could be some kind of dense matching. Afaik there is a french team which try to registrate images by graph-based-matching, but I neither remember the name, nor whether this worked well. One idea coould be to coarse-align the image with your technique and afterwards start a dense matching by optical flow. If you are certain to know which pixel are correctly matched (and those pixel are not within a small part of the image), you can compute another homography with ALL the points (instead of using RANSAC).

Upvotes: 1

Related Questions