0

is there anyone know how to convert the rotation and translation matrices computed from ORB-SLAM to a homography matrix? In addition, the results from ORB-SLAM were normalized, is there any way to scale to the pixel size? thank!

L.Cai
  • 1
  • 1
  • Homography does not take translation into account. so Homography will be ur (camera intrinsic matrix)*(relative rotation) * (inverse of camera intrinsic matrix). You should change the rotation matrix as coordinate axes of your reference point for rotation matrix and image might differ. If you are looking for more details. Kindly explain the question. – Garvita Tiwari Aug 12 '17 at 14:36
  • @GarvitaTiwari I think it is: for a generic scene, we can compute the homography from a purely rotational camera motion (as you have mentioned). From a planar scene, we can compute the homography using the rotation, translation and the knowledge about the plane (see [here](https://en.wikipedia.org/wiki/Homography_(computer_vision))). – Catree Aug 12 '17 at 16:28
  • @Catree yes homography ideally is calculated from rotation+translation+plane info. But L.Cai is using ORB-SLAM and opencv library I think, if he goes for further optimization using opencv library, translation has to be neglected. – Garvita Tiwari Aug 13 '17 at 03:39
  • @Garvita Tiwari I have experience on using the SIFT or SURF to extract the keypoints from 2 image then using opencv to calculate the homography. the scale is not changed but it is very slow. From orb-slam, the depth and all the points coordinate were normalized, so that the resulting camera rotation and translation parameters were also normalized. I hope to use the rotation and translation to calculate the homography then I can stitch the images. – L.Cai Aug 13 '17 at 06:46

0 Answers0