APORater.hpp

Go to the documentation of this file.

#pragma once

#include "BaseRater.hpp"

namespace ISM {
class APORater : public BaseRater
{

public:
    APORater(double bin_size, double maxAngleDeviation) : BaseRater(bin_size, maxAngleDeviation)
    {
      //We calculated a factor for sigma so that exp() is 0.5 at sqrt(3)*/(bin_size*2).
      // https://www.wolframalpha.com/input/?i=e%5E(((-(sqrt(3)%2F2*s)%5E2))%2F(2*(sigma)%5E2))%3D0.5
      squaredSigma = -0.735534 * bin_size * -0.735534 * bin_size;
    }

  //Methods are inlined for performance reasons.

  // No input checks at all since that method must run fast.
  virtual double rateAtBackProjectionLevel(VotedPosePtr& votedPose, RatingDataPtr data){

    /* Note: rate position with density function of a normal distribution without the scaling factor,
       so that the score is in [0,1] */
    double positionScore;
    positionScore = exp(-0.5 * (data->backProjectionSquaredDistance / squaredSigma));

    /* Note: rate orientation with a customized cosinus function, which resemble the density function above
       and the score is in [0,1]*/
    double orientationScore;

    //We doubled the limit for cos() so that this function is 0.5 at angleDeviation == maxAngleDeviation.
    // doubled area of angles with values > 0
    orientationScore = (data->angleDeviation < 2 * maxAngleDeviation) ? 0.5 + 0.5 * cos((M_PI * data->angleDeviation) / ( 2 * maxAngleDeviation)) : 0.0;

    double poseScore = positionScore * orientationScore;

    double totalScore = votedPose->weight - 1.0 + poseScore;

    /*Option to rate appearance in the furture*/
    //double appearanceScore = 1.0;
    //totalScore *= appearanceScore;

    return totalScore;
    
  }

  virtual void printRatingAtBackProjectionLevel(VotedPosePtr& votedPose, PosePtr& refPose, RatingDataPtr data){

    if (data == nullptr)
      data = RatingDataPtr(new RatingData());

    if (!this->isVoteSupportingReference(votedPose, refPose, data))
      {
        std::cout << "Doesn't support reference -> APORater score=0.0" << std::endl;
        return;
      }

    /* Note: rate position with density function of a normal distribution without the scaling factor,
       so that the score is in [0,1] */
    double positionScore;
    positionScore = exp(-0.5 * (data->backProjectionSquaredDistance / squaredSigma));

    /* Note: rate orientation with a customized cosinus function, which resemble the density function above
       and the score is in [0,1]*/
    double orientationScore;

    //We doubled the limit for cos() so that this function is 0.5 at angleDeviation == maxAngleDeviation.
    // doubled area of angles with values > 0
    orientationScore = (data->angleDeviation < 2 * maxAngleDeviation) ? 0.5 + 0.5 * cos((M_PI * data->angleDeviation) / ( 2 * maxAngleDeviation)) : 0.0;

    double poseScore = positionScore * orientationScore;

    double totalScore;
    if (votedPose->source->type.find("_sub") != std::string::npos)  //sub-ism
      {
        totalScore = votedPose->weight - 1.0 + poseScore;
      }
    else //object
      {
        totalScore = poseScore;
        /*Option to rate appearance in the furture*/
        //double appearanceScore = 1.0;
        //totalScore = appearanceScore * poseScore;
      }

    std::cout << "APORater score=" << totalScore << " with"
              << "\n\tPositionScore=" << positionScore << " with squaredDistance=" << data->backProjectionSquaredDistance << "  maxSquaredDistancepDeviation=" << (squaredSigma * 9) << "  squaredSigma=" << squaredSigma
              << "\n\tOrientationScore=" << orientationScore << " with angleDeviation=" << data->angleDeviation << "  maxAngleDeviation=" << maxAngleDeviation
              << "\nNormalized score=" << totalScore/votedPose->weightDenominator << " with denominator: " << votedPose->weightDenominator << std::endl;
 
    return;
    
  }

private:
    double squaredSigma;
};
}