Skip to content

O
opencv
Commit 2dc0cf23 authored by Maria Dimashova's avatar Maria Dimashova
Browse files
Options

added plot data generation for detectors; removed old version of functions evaluating detectors

parent 6667d935
Hide whitespace changes
Inline Side-by-side
Showing with 61 additions and 211 deletions
tests/cv/src/adetectordescriptor_evaluation.cpp
View file @ 2dc0cf23
......@@ -42,12 +42,15 @@
#include "cvtest.h"
#include <limits>
#include <cstdio>
#include <iostream>
#include <fstream>
using namespace std;
using namespace cv;
#define AFFINE_COVARIANT_VERSION
/****************************************************************************************\
* Functions to evaluate affine covariant detectors and descriptors. *
\****************************************************************************************/
inline Point2f applyHomography( const Mat_<double>& H, const Point2f& pt )
{
double z = H(2,0)*pt.x + H(2,1)*pt.y + H(2,2);
......@@ -78,123 +81,6 @@ inline void linearizeHomographyAt( const Mat_<double>& H, const Point2f& pt, Mat
A.setTo(Scalar::all(numeric_limits<double>::max()));
}
#ifndef AFFINE_COVARIANT_VERSION
/****************************************************************************************\
* 1. Initial version of evaluating detectors. This version calculate repeatability *
* for scale invariant detectors (circular regions) *
\****************************************************************************************/
// Find the key points located in the part of the scene present in both images
// and project keypoints2 on img1
void getCircularKeyPointsInCommonPart( const Mat& img1, const Mat img2, const Mat& H12,
const vector<KeyPoint>& keypoints1, const vector<KeyPoint>& keypoints2,
vector<KeyPoint>& ckeypoints1, vector<KeyPoint>& ckeypoints2t )
{
assert( !img1.empty() && !img2.empty() );
assert( !H12.empty() && H12.cols==3 && H12.rows==3 && H12.type()==CV_64FC1 );
ckeypoints1.clear();
ckeypoints2t.clear();
Rect r1(0, 0, img1.cols, img1.rows), r2(0, 0, img2.cols, img2.rows);
Mat H21; invert( H12, H21 );
for( vector<KeyPoint>::const_iterator it = keypoints1.begin();
it != keypoints1.end(); ++it )
{
if( r2.contains(applyHomography(H12, it->pt)) )
ckeypoints1.push_back(*it);
}
for( vector<KeyPoint>::const_iterator it = keypoints2.begin();
it != keypoints2.end(); ++it )
{
Point2f pt = applyHomography(H21, it->pt);
if( r1.contains(pt) )
{
KeyPoint kp = *it;
kp.pt = pt;
Mat_<double> A, eval;
linearizeHomographyAt(H21, it->pt, A);
eigen(A, eval);
assert( eval.type()==CV_64FC1 && eval.cols==1 && eval.rows==2 );
kp.size *= sqrt(eval(0,0) * eval(1,0)) /*scale from linearized homography matrix*/;
ckeypoints2t.push_back(kp);
}
}
}
// Locations p1 and p2 are repeated if ||p1 - H21*p2|| < 1.5 pixels.
// Regions are repeated if Es < 0.4 (Es differs for scale invariant and affine invarian detectors).
// For more details see "Scale&Affine Invariant Interest Point Detectors", Mikolajczyk, Schmid.
void evaluateScaleInvDetectors( const Mat& img1, const Mat img2, const Mat& H12,
const vector<KeyPoint>& keypoints1, const vector<KeyPoint>& keypoints2,
int& repeatingLocationCount, float& repeatingLocationRltv,
int& repeatingRegionCount, float& repeatingRegionRltv )
{
const double locThreshold = 1.5,
regThreshold = 0.4;
assert( !img1.empty() && !img2.empty() );
assert( !H12.empty() && H12.cols==3 && H12.rows==3 && H12.type()==CV_64FC1 );
Mat H21; invert( H12, H21 );
vector<KeyPoint> ckeypoints1, ckeypoints2t;
getCircularKeyPointsInCommonPart( img1, img2, H12, keypoints1, keypoints2, ckeypoints1, ckeypoints2t );
vector<KeyPoint> *smallKPSet = &ckeypoints1, *bigKPSet = &ckeypoints2t;
if( ckeypoints1.size() > ckeypoints2t.size() )
{
smallKPSet = &ckeypoints2t;
bigKPSet = &ckeypoints1;
}
if( smallKPSet->size() == 0 )
{
repeatingLocationCount = repeatingRegionCount = -1;
repeatingLocationRltv = repeatingRegionRltv = -1.f;
}
else
{
vector<bool> matchedMask( bigKPSet->size(), false);
repeatingLocationCount = repeatingRegionCount = 0;
for( vector<KeyPoint>::const_iterator skpIt = smallKPSet->begin(); skpIt != smallKPSet->end(); ++skpIt )
{
int nearestIdx = -1, bkpIdx = 0;
double minDist = numeric_limits<double>::max();
vector<KeyPoint>::const_iterator nearestBkp;
for( vector<KeyPoint>::const_iterator bkpIt = bigKPSet->begin(); bkpIt != bigKPSet->end(); ++bkpIt, bkpIdx++ )
{
if( !matchedMask[bkpIdx] )
{
Point p1(cvRound(skpIt->pt.x), cvRound(skpIt->pt.y)),
p2(cvRound(bkpIt->pt.x), cvRound(bkpIt->pt.y));
double dist = norm(p1 - p2);
if( dist < minDist )
{
nearestIdx = bkpIdx;
minDist = dist;
nearestBkp = bkpIt;
}
}
}
if( minDist < locThreshold )
{
matchedMask[nearestIdx] = true;
repeatingLocationCount++;
double minRadius = min( skpIt->size, nearestBkp->size ),
maxRadius = max( skpIt->size, nearestBkp->size );
double Es = abs(1 - (minRadius*minRadius)/(maxRadius*maxRadius));
if( Es < regThreshold )
repeatingRegionCount++;
}
}
repeatingLocationRltv = smallKPSet->size() ? (float)repeatingLocationCount / smallKPSet->size() : 0;
repeatingRegionRltv = smallKPSet->size() ? (float)repeatingRegionCount / smallKPSet->size() : 0;
}
}
#else
/****************************************************************************************\
* 2. Functions to evaluate affine covariant detectors and descriptors. *
\****************************************************************************************/
class EllipticKeyPoint
{
public:
......@@ -486,45 +372,35 @@ void calculateRepeatability( const vector<EllipticKeyPoint>& _keypoints1, const
if( !size || overlaps.nzcount() == 0 )
return;
// threshold the overlaps
for( int y = 0; y < size[0]; y++ )
{
for( int x = 0; x < size[1]; x++ )
{
if ( overlaps(y,x) < overlapThreshold )
overlaps.erase(y,x);
}
}
if( ifEvaluateDetectors )
{
// regions one-to-one matching
correspondencesCount = 0;
SparseMat_<float> currOverlaps( 2, size );
for( int y = 0; y < size[0]; y++ )
{
for( int x = 0; x < size[1]; x++ )
{
float val = overlaps(y,x);
if ( val >= overlapThreshold )
currOverlaps.ref(y,x) = val;
}
}
while( currOverlaps.nzcount() > 0 )
while( overlaps.nzcount() > 0 )
{
double maxOverlap = 0;
int maxIdx[2];
minMaxLoc( currOverlaps, 0, &maxOverlap, 0, maxIdx );
minMaxLoc( overlaps, 0, &maxOverlap, 0, maxIdx );
for( size_t i1 = 0; i1 < keypoints1.size(); i1++ )
currOverlaps.erase(i1, maxIdx[1]);
overlaps.erase(i1, maxIdx[1]);
for( size_t i2 = 0; i2 < keypoints2t.size(); i2++ )
currOverlaps.erase(maxIdx[0], i2);
overlaps.erase(maxIdx[0], i2);
correspondencesCount++;
}
repeatability = minCount ? (float)(correspondencesCount*100)/minCount : -1;
}
else
{
thresholdedOverlapMask->create( 2, size );
for( int y = 0; y < size[0]; y++ )
{
for( int x = 0; x < size[1]; x++ )
{
float val = overlaps(y,x);
if ( val >= overlapThreshold )
thresholdedOverlapMask->ref(y,x) = val;
}
}
overlaps.copyTo(*thresholdedOverlapMask);
}
}
......@@ -588,7 +464,6 @@ void evaluateDescriptors( const vector<EllipticKeyPoint>& keypoints1, const vect
}
}
#endif
/****************************************************************************************\
* Detectors evaluation *
\****************************************************************************************/
......@@ -603,15 +478,8 @@ const string KEYPOINTS_DIR = "detectors_descriptors_evaluation/keypoints_dataset
const string PARAMS_POSTFIX = "_params.xml";
const string RES_POSTFIX = "_res.xml";
#ifndef AFFINE_COVARIANT_VERSION
const string RLC = "repeating_locations_count";
const string RLR = "repeating_locations_rltv";
const string RRC = "repeating_regions_count";
const string RRR = "repeating_regions_rltv";
#else
const string REPEAT = "repeatability";
const string CORRESP_COUNT = "correspondence_count";
#endif
string DATASET_NAMES[DATASETS_COUNT] = { "bark", "bikes", "boat", "graf", "leuven", "trees", "ubc", "wall"};
......@@ -666,7 +534,7 @@ protected:
virtual void processResults();
virtual int processResults( int datasetIdx, int caseIdx ) = 0;
void writeAllPlotData() const;
virtual void writePlotData( const string &filename, int datasetIdx ) const {};
virtual void writePlotData( int datasetIdx ) const {};
string algName;
bool isWriteParams, isWriteResults, isWriteGraphicsData;
......@@ -861,11 +729,7 @@ void BaseQualityTest::writeAllPlotData() const
{
for( int di = 0; di < DATASETS_COUNT; di++ )
{
stringstream stream;
stream << getPlotPath() << algName << "_" << DATASET_NAMES[di] << ".csv";
string filename;
stream >> filename;
writePlotData( filename, di );
writePlotData( di );
}
}
......@@ -952,9 +816,8 @@ protected:
virtual void readDefaultRunParams( FileNode &fn );
virtual void writeDefaultRunParams( FileStorage &fs ) const;
virtual void writePlotData( int di ) const;
Ptr<FeatureDetector> specificDetector;
Ptr<FeatureDetector> defaultDetector;
void openToWriteKeypointsFile( FileStorage& fs, int datasetIdx );
virtual void readAlgorithm( );
......@@ -962,20 +825,17 @@ protected:
virtual void runDatasetTest( const vector<Mat> &imgs, const vector<Mat> &Hs, int di, int &progress );
virtual int processResults( int datasetIdx, int caseIdx );
Ptr<FeatureDetector> specificDetector;
Ptr<FeatureDetector> defaultDetector;
struct Quality
{
#ifndef AFFINE_COVARIANT_VERSION
int repeatingLocationCount;
float repeatingLocationRltv;
int repeatingRegionCount;
float repeatingRegionRltv;
#else
float repeatability;
int correspondenceCount;
#endif
};
vector<vector<Quality> > validQuality;
vector<vector<Quality> > calcQuality;
vector<bool> isSaveKeypoints;
vector<bool> isActiveParams;
......@@ -1025,28 +885,14 @@ bool DetectorQualityTest::isCalcQualityEmpty( int datasetIdx ) const
void DetectorQualityTest::readResults( FileNode& fn, int datasetIdx, int caseIdx )
{
#ifndef AFFINE_COVARIANT_VERSION
validQuality[datasetIdx][caseIdx].repeatingLocationCount = fn[RLC];
validQuality[datasetIdx][caseIdx].repeatingLocationRltv = fn[RLR];
validQuality[datasetIdx][caseIdx].repeatingRegionCount = fn[RRC];
validQuality[datasetIdx][caseIdx].repeatingRegionRltv = fn[RRR];
#else
validQuality[datasetIdx][caseIdx].repeatability = fn[REPEAT];
validQuality[datasetIdx][caseIdx].correspondenceCount = fn[CORRESP_COUNT];
#endif
}
void DetectorQualityTest::writeResults( FileStorage& fs, int datasetIdx, int caseIdx ) const
{
#ifndef AFFINE_COVARIANT_VERSION
fs << RLC << calcQuality[datasetIdx][caseIdx].repeatingLocationCount;
fs << RLR << calcQuality[datasetIdx][caseIdx].repeatingLocationRltv;
fs << RRC << calcQuality[datasetIdx][caseIdx].repeatingRegionCount;
fs << RRR << calcQuality[datasetIdx][caseIdx].repeatingRegionRltv;
#else
fs << REPEAT << calcQuality[datasetIdx][caseIdx].repeatability;
fs << CORRESP_COUNT << calcQuality[datasetIdx][caseIdx].correspondenceCount;
#endif
}
void DetectorQualityTest::readDefaultRunParams (FileNode &fn)
......@@ -1091,6 +937,36 @@ void DetectorQualityTest::setDefaultDatasetRunParams( int datasetIdx )
isActiveParams[datasetIdx] = isActiveParamsDefault;
}
void DetectorQualityTest::writePlotData(int di ) const
{
int imgXVals[] = { 2, 3, 4, 5, 6 }; // if scale, blur or light changes
int viewpointXVals[] = { 20, 30, 40, 50, 60 }; // if viewpoint changes
int jpegXVals[] = { 60, 80, 90, 95, 98 }; // if jpeg compression
int* xVals = 0;
if( !DATASET_NAMES[di].compare("ubc") )
{
xVals = jpegXVals;
}
else if( !DATASET_NAMES[di].compare("graf") || !DATASET_NAMES[di].compare("wall") )
{
xVals = viewpointXVals;
}
else
xVals = imgXVals;
stringstream rFilename, cFilename;
rFilename << getPlotPath() << algName << "_" << DATASET_NAMES[di] << "_repeatability.csv";
cFilename << getPlotPath() << algName << "_" << DATASET_NAMES[di] << "_correspondenceCount.csv";
ofstream rfile(rFilename.str().c_str()), cfile(cFilename.str().c_str());
for( int ci = 0; ci < TEST_CASE_COUNT; ci++ )
{
rfile << xVals[ci] << ", " << calcQuality[di][ci].repeatability << endl;
cfile << xVals[ci] << ", " << calcQuality[di][ci].correspondenceCount << endl;
}
}
void DetectorQualityTest::openToWriteKeypointsFile( FileStorage& fs, int datasetIdx )
{
string filename = string(ts->get_data_path()) + KEYPOINTS_DIR + algName + "_"+
......@@ -1194,16 +1070,10 @@ void DetectorQualityTest::runDatasetTest (const vector<Mat> &imgs, const vector<
vector<KeyPoint> keypoints2;
detector->detect( imgs[ci+1], keypoints2 );
writeKeypoints( keypontsFS, keypoints2, ci+1);
#ifndef AFFINE_COVARIANT_VERSION
evaluateScaleInvDetectors( imgs[0], imgs[ci+1], Hs[ci], keypoints1, keypoints2,
calcQuality[di][ci].repeatingLocationCount, calcQuality[di][ci].repeatingLocationRltv,
calcQuality[di][ci].repeatingRegionCount, calcQuality[di][ci].repeatingRegionRltv );
#else
vector<EllipticKeyPoint> ekeypoints2;
transformToEllipticKeyPoints( keypoints2, ekeypoints2 );
evaluateDetectors( ekeypoints1, ekeypoints2, imgs[0], imgs[ci], Hs[ci],
calcQuality[di][ci].repeatability, calcQuality[di][ci].correspondenceCount );
#endif
}
}
......@@ -1224,27 +1094,6 @@ int DetectorQualityTest::processResults( int datasetIdx, int caseIdx )
bool isBadAccuracy;
int countEps = 1;
const float rltvEps = 0.001;
#ifndef AFFINE_COVARIANT_VERSION
ts->printf(CvTS::LOG, "%s: calc=%d, valid=%d", RLC.c_str(), calc.repeatingLocationCount, valid.repeatingLocationCount );
isBadAccuracy = valid.repeatingLocationCount - calc.repeatingLocationCount > countEps;
testLog( ts, isBadAccuracy );
res = isBadAccuracy ? CvTS::FAIL_BAD_ACCURACY : res;
ts->printf(CvTS::LOG, "%s: calc=%f, valid=%f", RLR.c_str(), calc.repeatingLocationRltv, valid.repeatingLocationRltv );
isBadAccuracy = valid.repeatingLocationRltv - calc.repeatingLocationRltv > rltvEps;
testLog( ts, isBadAccuracy );
res = isBadAccuracy ? CvTS::FAIL_BAD_ACCURACY : res;
ts->printf(CvTS::LOG, "%s: calc=%d, valid=%d", RRC.c_str(), calc.repeatingRegionCount, valid.repeatingRegionCount );
isBadAccuracy = valid.repeatingRegionCount - calc.repeatingRegionCount > countEps;
testLog( ts, isBadAccuracy );
res = isBadAccuracy ? CvTS::FAIL_BAD_ACCURACY : res;
ts->printf(CvTS::LOG, "%s: calc=%f, valid=%f", RRR.c_str(), calc.repeatingRegionRltv, valid.repeatingRegionRltv );
isBadAccuracy = valid.repeatingRegionRltv - calc.repeatingRegionRltv > rltvEps;
testLog( ts, isBadAccuracy );
res = isBadAccuracy ? CvTS::FAIL_BAD_ACCURACY : res;
#else
ts->printf(CvTS::LOG, "%s: calc=%f, valid=%f", REPEAT.c_str(), calc.repeatability, valid.repeatability );
isBadAccuracy = valid.repeatability - calc.repeatability > rltvEps;
testLog( ts, isBadAccuracy );
......@@ -1254,7 +1103,6 @@ int DetectorQualityTest::processResults( int datasetIdx, int caseIdx )
isBadAccuracy = valid.correspondenceCount - calc.correspondenceCount > countEps;
testLog( ts, isBadAccuracy );
res = isBadAccuracy ? CvTS::FAIL_BAD_ACCURACY : res;
#endif
return res;
}
......@@ -1328,7 +1176,7 @@ protected:
virtual int processResults( int datasetIdx, int caseIdx );
virtual void writePlotData( const string &filename, int di ) const;
virtual void writePlotData( int di ) const;
struct Quality
{
......@@ -1455,9 +1303,11 @@ void DescriptorQualityTest::setDefaultDatasetRunParams( int datasetIdx )
commRunParams[datasetIdx].keypontsFilename = "surf_" + DATASET_NAMES[datasetIdx] + ".xml.gz";
}
void DescriptorQualityTest::writePlotData( const string &filename, int di ) const
void DescriptorQualityTest::writePlotData( int di ) const
{
FILE *file = fopen (filename.c_str(),"w");
stringstream filename;
filename << getPlotPath() << algName << "_" << DATASET_NAMES[di] << ".csv";
FILE *file = fopen (filename.str().c_str(), "w");
size_t size = calcDatasetQuality[di].size();
for (size_t i=0;i<size;i++)
{
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment