model_search_impl.h

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2019, Individual contributors, see AUTHORS file     |
   | See: https://www.mrpt.org/Authors - All rights reserved.               |
   | Released under BSD License. See: https://www.mrpt.org/License          |
   +------------------------------------------------------------------------+ */

#pragma once

#ifndef math_modelsearch_h
#include "model_search.h"
#endif

#include <cmath>
#include <limits>

namespace mrpt::math
{
//----------------------------------------------------------------------
//! Run the ransac algorithm searching for a single model
template <typename TModelFit>
bool ModelSearch::ransacSingleModel(
    const TModelFit& p_state, size_t p_kernelSize,
    const typename TModelFit::Real& p_fitnessThreshold,
    typename TModelFit::Model& p_bestModel, std::vector<size_t>& p_inliers)
{
    size_t bestScore = std::string::npos;  // npos will mean "none"
    size_t iter = 0;
    size_t softIterLimit = 1;  // will be updated by the size of inliers
    size_t hardIterLimit = 100;  // a fixed iteration step
    p_inliers.clear();
    size_t nSamples = p_state.getSampleCount();
    std::vector<size_t> ind(p_kernelSize);

    while (iter < softIterLimit && iter < hardIterLimit)
    {
        bool degenerate = true;
        typename TModelFit::Model currentModel;
        size_t i = 0;
        while (degenerate)
        {
            pickRandomIndex(nSamples, p_kernelSize, ind);
            degenerate = !p_state.fitModel(ind, currentModel);
            i++;
            if (i > 100) return false;
        }

        std::vector<size_t> inliers;

        for (size_t j = 0; j < nSamples; j++)
        {
            if (p_state.testSample(j, currentModel) < p_fitnessThreshold)
                inliers.push_back(j);
        }
        ASSERT_(inliers.size() > 0);

        // Find the number of inliers to this model.
        const size_t ninliers = inliers.size();
        bool update_estim_num_iters =
            (iter == 0);  // Always update on the first iteration, regardless of
        // the result (even for ninliers=0)

        if (ninliers > bestScore ||
            (bestScore == std::string::npos && ninliers != 0))
        {
            bestScore = ninliers;
            p_bestModel = currentModel;
            p_inliers = inliers;
            update_estim_num_iters = true;
        }

        if (update_estim_num_iters)
        {
            // Update the estimation of maxIter to pick dataset with no outliers
            // at propability p
            double f = ninliers / static_cast<double>(nSamples);
            double p = 1 - std::pow(f, static_cast<double>(p_kernelSize));
            const double eps = std::numeric_limits<double>::epsilon();
            p = std::max(eps, p);  // Avoid division by -Inf
            p = std::min(1 - eps, p);  // Avoid division by 0.
            softIterLimit = log(1 - p) / log(p);
        }

        iter++;
    }

    return true;
}

//----------------------------------------------------------------------
//! Run a generic programming version of ransac searching for a single model
template <typename TModelFit>
bool ModelSearch::geneticSingleModel(
    const TModelFit& p_state, size_t p_kernelSize,
    const typename TModelFit::Real& p_fitnessThreshold, size_t p_populationSize,
    size_t p_maxIteration, typename TModelFit::Model& p_bestModel,
    std::vector<size_t>& p_inliers)
{
    // a single specie of the population
    using Species = TSpecies<TModelFit>;
    std::vector<Species> storage;
    std::vector<Species*> population;
    std::vector<Species*> sortedPopulation;

    size_t sampleCount = p_state.getSampleCount();
    int elderCnt = (int)p_populationSize / 3;
    int siblingCnt = (p_populationSize - elderCnt) / 2;
    int speciesAlive = 0;

    storage.resize(p_populationSize);
    population.reserve(p_populationSize);
    sortedPopulation.reserve(p_populationSize);
    for (typename std::vector<Species>::iterator it = storage.begin();
         it != storage.end(); it++)
    {
        pickRandomIndex(sampleCount, p_kernelSize, it->sample);
        population.push_back(&*it);
        sortedPopulation.push_back(&*it);
    }

    size_t iter = 0;
    while (iter < p_maxIteration)
    {
        if (iter > 0)
        {
            // generate new population: old, siblings,  new
            population.clear();
            int i = 0;

            // copy the best elders
            for (; i < elderCnt; i++)
            {
                population.push_back(sortedPopulation[i]);
            }

            // mate elders to make siblings
            int se = (int)speciesAlive;  // dead species cannot mate
            for (; i < elderCnt + siblingCnt; i++)
            {
                Species* sibling = sortedPopulation[--se];
                population.push_back(sibling);

                // pick two parents, from the species not yet refactored
                // better elders has more chance to mate as they are removed
                // later from the list
                int r1 = rand();
                int r2 = rand();
                int p1 = r1 % se;
                int p2 =
                    (p1 > se / 2) ? (r2 % p1) : p1 + 1 + (r2 % (se - p1 - 1));
                ASSERT_(p1 != p2 && p1 < se && p2 < se);
                ASSERT_(se >= elderCnt);
                Species* a = sortedPopulation[p1];
                Species* b = sortedPopulation[p2];

                // merge the sample candidates
                std::set<size_t> sampleSet;
                sampleSet.insert(a->sample.begin(), a->sample.end());
                sampleSet.insert(b->sample.begin(), b->sample.end());
                // mutate - add a random sample that will be selected with some
                // (non-zero) probability
                sampleSet.insert(rand() % sampleCount);
                pickRandomIndex(sampleSet, p_kernelSize, sibling->sample);
            }

            // generate some new random species
            for (; i < (int)p_populationSize; i++)
            {
                Species* s = sortedPopulation[i];
                population.push_back(s);
                pickRandomIndex(sampleCount, p_kernelSize, s->sample);
            }
        }

        // evaluate species
        speciesAlive = 0;
        for (typename std::vector<Species*>::iterator it = population.begin();
             it != population.end(); it++)
        {
            Species& s = **it;
            if (p_state.fitModel(s.sample, s.model))
            {
                s.fitness = 0;
                for (size_t i = 0; i < p_state.getSampleCount(); i++)
                {
                    typename TModelFit::Real f = p_state.testSample(i, s.model);
                    if (f < p_fitnessThreshold)
                    {
                        s.fitness += f;
                        s.inliers.push_back(i);
                    }
                }
                ASSERT_(s.inliers.size() > 0);

                s.fitness /= s.inliers.size();
                // scale by the number of outliers
                s.fitness *= (sampleCount - s.inliers.size());
                speciesAlive++;
            }
            else
                s.fitness =
                    std::numeric_limits<typename TModelFit::Real>::max();
        }

        if (!speciesAlive)
        {
            // the world is dead, no model was found
            return false;
        }

        // sort by fitness (ascending)
        std::sort(
            sortedPopulation.begin(), sortedPopulation.end(), Species::compare);

        iter++;
    }

    p_bestModel = sortedPopulation[0]->model;
    p_inliers = sortedPopulation[0]->inliers;

    return !p_inliers.empty();
}
}  // namespace mrpt::math