TrainEdgePottsCS.cpp

#include "TrainEdgePottsCS.h"
#include "macroses.h"

namespace DirectGraphicalModels
{
// Calculates the Euclidean distance betwee two feature vectors
float calculateContrast(const Mat &featureVector1, const Mat &featureVector2)
{
    Mat     fv1, fv2, dfv;
    float   res = 0.0f;
    int nFeatures = featureVector1.rows;
    
    featureVector1.convertTo(fv1, CV_32FC1);
    featureVector2.convertTo(fv2, CV_32FC1);

    // distance between feature vectors
    subtract(fv1, fv2, dfv);                // dfv = fv1 - fv2;
    multiply(dfv, dfv, dfv);                // sqr(dfv);
    for (int i = 0; i < nFeatures; i++) 
        res += dfv.at<float>(i, 0);
    res = sqrtf(res / nFeatures);

    fv1.release();
    fv2.release();
    dfv.release();  

    return res;
}

// Charbonnier Penalizer
float penalizerChar(float x, float l)
{
    float res = l / sqrt(l*l + x*x);
    return MAX(FLT_EPSILON, res);
}

// Perona - Malik Penaliter
float penalizerPM(float x, float l)
{
    float res = l*l / (l*l + x*x);
    return MAX(FLT_EPSILON, res);
}

// Exponential penalizer
float penalizerExp (float x, float l)
{
    float res = expf(-l * x*x);         
    return MAX(FLT_EPSILON, res);
}


Mat CTrainEdgePottsCS::calculateEdgePotentials(const Mat &featureVector1, const Mat &featureVector2, const vec_float_t &vParams) const
{
    DGM_ASSERT_MSG((vParams.size() == 2) || (vParams.size() == m_nStates + 1), "Wrong number of parameters: %zu. It must be either %d or %u", vParams.size(), 2, m_nStates + 1);

    Mat res = CTrainEdgePotts::calculateEdgePotentials(featureVector1, featureVector2, vec_float_t(vParams.begin(), vParams.end() - 1));
    if (featureVector1.empty() || featureVector2.empty()) return res;   // no cotrast could be calcilated -> return potts edge potential

    // Assertions:
    DGM_ASSERT_MSG((featureVector1.type() == CV_8UC1) && (featureVector2.type() == CV_8UC1), 
        "One (or both) of input feature vectors has either wrong depth or more than one channel");
    DGM_ASSERT_MSG((featureVector1.size().width == 1) && (featureVector1.size().height == getNumFeatures()), 
        "The first input feature vector has wrong size:(%d, %d)", featureVector1.size().width, featureVector1.size().height);
    DGM_ASSERT_MSG((featureVector2.size().width == 1) && (featureVector2.size().height == getNumFeatures()),
        "The second input feature vector has wrong size:(%d, %d)", featureVector2.size().width, featureVector2.size().height);

    float penalty;
    float dst = calculateContrast(featureVector1, featureVector2);
    switch(m_penApproach) {
        case eP_APP_PEN_CHAR:   penalty = penalizerChar(dst, vParams.back());   break;
        case eP_APP_PEN_PM:     penalty = penalizerPM(dst, vParams.back());     break;
        case eP_APP_PEN_EXP:    penalty = penalizerExp(dst, vParams.back());    break;
    }

    for (byte s = 0; s < m_nStates; s++) res.at<float>(s, s) = MAX(1.0f, res.at<float>(s, s) * penalty);

    return res;
}
}