Function SolveProblem

examples/denoising.cc:211–251

Solves the FoE problem using Ceres and post-processes it to make sure the solution stays within [0, 255].

// Solves the FoE problem using Ceres and post-processes it to make sure the
// solution stays within [0, 255].
void SolveProblem(Problem* problem, PGMImage<double>* solution) {
  // These parameters may be experimented with. For example, ceres::DOGLEG tends
  // to be faster for 2x2 filters, but gives solutions with slightly higher
  // objective function value.
  ceres::Solver::Options options;
  SetMinimizerOptions(&options);
  SetLinearSolver(&options);
  options.function_tolerance = 1e-3;  // Enough for denoising.

  if (options.linear_solver_type == ceres::CGNR &&
      options.preconditioner_type == ceres::SUBSET) {
    std::vector<ResidualBlockId> residual_blocks;
    problem->GetResidualBlocks(&residual_blocks);

    // To use the SUBSET preconditioner we need to provide a list of
    // residual blocks (rows of the Jacobian). The denoising problem
    // has fairly general sparsity, and there is no apriori reason to
    // select one residual block over another, so we will randomly
    // subsample the residual blocks with probability subset_fraction.
    std::default_random_engine engine;
    std::uniform_real_distribution<> distribution(0, 1);  // rage 0 - 1
    for (auto residual_block : residual_blocks) {
      if (distribution(engine) <= CERES_GET_FLAG(FLAGS_subset_fraction)) {
        options.residual_blocks_for_subset_preconditioner.insert(
            residual_block);
      }
    }
  }

  ceres::Solver::Summary summary;
  ceres::Solve(options, problem, &summary);
  std::cout << summary.FullReport() << "\n";

  // Make the solution stay in [0, 255].
  for (int x = 0; x < solution->width(); ++x) {
    for (int y = 0; y < solution->height(); ++y) {
      *solution->MutablePixel(x, y) =
          std::min(255.0, std::max(0.0, solution->Pixel(x, y)));
    }
  }
}

}  // namespace
}  // namespace ceres::examples