0

I am implementing my own image stitching algorithms(using opencv but not the stitcher class), so far i can stitch 2 images using following steps:

  1. Detect keypoints
  2. Descript keypoints
  3. Match keypoints
  4. Calculate homography
  5. Warp images
  6. Blend images

What i'd like to know is, that if let's say i'd like to do video stitching which means i need this algorithm to be fast, could i skip the steps 1-4? i would only compute homography once and then use same matrix for all other frames - of course cameras would be in static position

mereth
  • 465
  • 3
  • 9
  • 19
  • Yes. Also depends on stuff like focus, but basically if you know what and where to rotate - you don't need 1-4 steps. – Dmitrii Z. Feb 27 '18 at 22:45
  • That is basically what i wanted, if you want you can copy it as an answer and i'll close this – mereth Feb 28 '18 at 10:50

1 Answers1

0

Yes you can do that.

You should keep track of different camera parameters like focus which change homography, but basically if you know what and where to rotate - you can skip steps 1-4.

You can see stitching_detail.cpp from opencv source code which despite being buggy in camera parameters estimation and a little bit memory & CPU inefficient, greatly demonstrates stitching pipeline.

Dmitrii Z.
  • 2,287
  • 3
  • 19
  • 29