In this homework, we are again given 2D and 3D points and their correspondence. On top of that, we assume the camera is calibrated and the calibration matrix is given. There are three steps before we finalize our answer to this problem: outlier rejection, linear estimation, and non-linear estimation. For outlier rejection, its maximum number of trials based on an algorithm called M-estimator Sample Consensus (mSAC) algorithm. For linear estimation, we use a framework called Efficient Perspective-n-Point (EPnP) algorithm to find out the camera pose. After that, we improve the result from linear estimation with Levenberg-Marquardt (LM) algorithm. After this, the Root Mean Squared Error (RMSE) drops significantly from 1.474598 to 1.215311.
[pdf_embed url=’https://hogansung.com/wp-content/uploads/2021/02/graduate-Winter2016-ComputerVision_II_Hw3-report.pdf’