Revision history [back]

from https://pointclouds.org/documentation/classpcl_1_1_random_sample_consensus.html

The algorithm works as follows:

1. randomly select samples from the cloud, just as many as needed to determine a model
2. compute the coefficients of the model from the samples
3. count how many points of the cloud belong to the model, given a threshold. These are called inliers
4. repeat until a good model has been found or a max number of iterations has been reached
5. return the model with the most inliers

I believe the answer to you question is related to stage 4, in which the model parameters, if setOptimizeCoefficients is true, is refined to fit all inliers, instead of using the original parameters that were fitted in stage 2 for the random samples only.