I have been using rrel_ran_sam_search class for finding 2D homographies.
I came across the following note in the code
00105 // The main loop repeatedly establishes a sample, generates fit
00106 // parameters from the sample, calculates the objective function
00107 // value, and tests the value against the best found thus far. If
00108 // a sample doesn't yield a parameter vector, it is still counted
00109 // toward the total number to take. This prevents errors that
00110 // would arise when all samples are to be used, but still works
00111 // correctly for probabilistic sampling because the possibility
00112 // is rare.
Instead of counting a sample that doesn't yield a parameter vector towards
Number to take , why not make an exception for probabilistic sampling?
By adding the following line
if ( problem->fit_from_minimal_set( point_indices, new_params ))
//here goes what to do if the sample generates a parameter set
//Note s is the number of sample to be taken
if (trace_level_ >= 1)
vcl_cout << "No fit to sample.\n";