3
$\begingroup$

I have two cameras side by side. I'd like to register two images taken from each camera. I think global transformation will not work for this issue since changes of distances between two images for the closer objects are significantly higher. What should I do in this case? I tried to read some papers but I am not sure which one is the perfect match for me. Do you have any suggestions such as "read this paper", "apply this algorithm", "know this and that" etc.

edit: I forgot to say that the project is real-time. Image registration will be implemented in the GPU.

| cite | improve this question | |
$\endgroup$
  • 2
    $\begingroup$ Good question! I hope someone can help you. Have you tried SIFT/SURF/global alignment? You're probably right that it will have some issues, but it's a reasonable thing to try first. See also cs.stackexchange.com/q/47234/755, though it doesn't answer this question (and this isn't a duplicate). $\endgroup$ – D.W.♦ Nov 21 '16 at 17:38
  • $\begingroup$ I actually read this one before asking this question. The answer is quite helpful but general. By the way, I'll try. Thank you. $\endgroup$ – Mustafa Işık Nov 21 '16 at 19:08
  • $\begingroup$ You should read about stereo image/stereo vision. Basically you can not register the images without building a 3D model. If the cameras` relative location is fixed it should not be too difficult (although it is definitely not trivial) and it is called calibrated stereo. Otherwise it becomes a bit more tricky and is called uncalibrated stereo $\endgroup$ – Rosa Gronchi Jan 30 '17 at 14:20
1
$\begingroup$

Since you need to implement image registration for real-time your should have fast registration algorithm, such as differential total variation (DTV) or binary gradient angle descriptor. However, these methods implement for challenge dataset or multimodal in consideration of your images are mono-modal meanwhile these methods should work perfectly in your case.

Your algorithm should include these steps:

  1. Image descriptor to extract the images features.
  2. Feature selection: find the most similar features (this step is important to remove outliers which is common)
  3. Minimise the distance between the related image features to get the optimal transformation.
  4. Apply image transformation on the floating image.
| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.