Non-Linear Refinement

The DLT algorithm is linear and fast, but it minimizes an algebraic error that has no direct physical meaning. For maximum precision, we must minimize the Geometric Reprojection Error.

Transfer Error

The goal is to find $H$ that minimizes the distance between the measured points and the projected points:

$\sum_i d(\mathbf{x}_i', H \mathbf{x}_i)^2 + d(\mathbf{x}_i, H^{-1} \mathbf{x}_i')^2$

This is a non-linear least squares problem. Since we already have a good initial guess from RANSAC/DLT, we can use iterative optimization.

Levenberg-Marquardt

The LM algorithm interpolates between the Gauss-Newton algorithm and the method of gradient descent. It is the gold standard for refining camera poses and homographies in professional vision pipelines.

The Final Step: Modern VPS systems always end with LM refinement. It squashes the last remaining sub-pixel errors, ensuring that the AR overlays or sensor alignments are pixel-perfect.

The Convergence Plot

Initial Error
5.82 px
Final Error
--
Iterations
0
Convergence
Idle