Camera pose estimation.

It is one of the most oldest problems in computer vision. Nowadays, there are many successful implements out there. In OpenCV alone, there are findFundamentalMat which can use 8 and 7 points method (also with RANSAC or LMEDS), and in opencv 3.x, there is an additional findEssentialMat which uses 5-points algorithm (also with RANSAC or LMEDS) which also have recoverPose to determine which of the four possible solution is really good.

However, in my experiments, I encountered some problems:

  • The E from findEssentialMat dose not equals to $K.T * F * K$
  • Sometimes it choose the wrong potential answer in recoverPose

So, I want to do extra experiments with different implements of 5-points and see how they really behaves. The candidates are:

UPDATE: 2018.11.26: Add pose_definition for the definition of POSE, R, t, C used in OpenCV.

Some Notes

Time of experiment

20170818, the last comment of Nghia Ho is APRIL 21, 2016 AT 8:11 PM


Good Luck & Have Fun.