Skip to content

O
opencv
Commit fcd999ae authored by Maria Dimashova's avatar Maria Dimashova
Browse files
Options

added SL2 (squared L2 distance) and implemented the descriptors matching in L2 using SL2

parent a9fdc1bd
Hide whitespace changes
Inline Side-by-side
Showing with 145 additions and 74 deletions
modules/features2d/doc/common_interfaces_of_descriptor_matchers.rst
View file @ fcd999ae
......@@ -300,6 +300,20 @@ For efficiency, ``BruteForceMatcher`` is used as a template parameterized with t
ResultType operator()( const T* a, const T* b, int size ) const;
};
/*
* Squared Euclidean distance functor
*/
template<class T>
struct SL2
{
typedef T ValueType;
typedef typename Accumulator<T>::Type ResultType;
ResultType operator()( const T* a, const T* b, int size ) const;
};
// Note: in case of SL2 distance a parameter maxDistance in the method DescriptorMatcher::radiusMatch
// is a squared maximum distance in L2.
/*
* Manhattan distance (city block distance) functor
......@@ -311,7 +325,6 @@ For efficiency, ``BruteForceMatcher`` is used as a template parameterized with t
typedef typename Accumulator<T>::Type ResultType;
ResultType operator()( const T* a, const T* b, int size ) const;
...
};
/*
......@@ -334,7 +347,6 @@ For efficiency, ``BruteForceMatcher`` is used as a template parameterized with t
ResultType operator()( const unsigned char* a, const unsigned char* b,
int size ) const;
...
};
......
modules/features2d/include/opencv2/features2d/features2d.hpp
View file @ fcd999ae
......@@ -2083,6 +2083,27 @@ template<> struct Accumulator<unsigned short> { typedef float Type; };
template<> struct Accumulator<char> { typedef float Type; };
template<> struct Accumulator<short> { typedef float Type; };
/*
* Squeared Euclidean distance functor
*/
template<class T>
struct CV_EXPORTS SL2
{
typedef T ValueType;
typedef typename Accumulator<T>::Type ResultType;
ResultType operator()( const T* a, const T* b, int size ) const
{
ResultType result = ResultType();
for( int i = 0; i < size; i++ )
{
ResultType diff = (ResultType)(a[i] - b[i]);
result += diff*diff;
}
return result;
}
};
/*
* Euclidean distance functor
*/
......@@ -2395,77 +2416,77 @@ template<class Distance>
inline void BruteForceMatcher<Distance>::commonKnnMatchImpl( BruteForceMatcher<Distance>& matcher,
const Mat& queryDescriptors, vector<vector<DMatch> >& matches, int knn,
const vector<Mat>& masks, bool compactResult )
{
typedef typename Distance::ValueType ValueType;
typedef typename Distance::ResultType DistanceType;
CV_DbgAssert( !queryDescriptors.empty() );
CV_Assert( DataType<ValueType>::type == queryDescriptors.type() );
{
typedef typename Distance::ValueType ValueType;
typedef typename Distance::ResultType DistanceType;
CV_DbgAssert( !queryDescriptors.empty() );
CV_Assert( DataType<ValueType>::type == queryDescriptors.type() );
int dimension = queryDescriptors.cols;
matches.reserve(queryDescriptors.rows);
int dimension = queryDescriptors.cols;
matches.reserve(queryDescriptors.rows);
size_t imgCount = matcher.trainDescCollection.size();
vector<Mat> allDists( imgCount ); // distances between one query descriptor and all train descriptors
for( size_t i = 0; i < imgCount; i++ )
size_t imgCount = matcher.trainDescCollection.size();
vector<Mat> allDists( imgCount ); // distances between one query descriptor and all train descriptors
for( size_t i = 0; i < imgCount; i++ )
allDists[i] = Mat( 1, matcher.trainDescCollection[i].rows, DataType<DistanceType>::type );
for( int qIdx = 0; qIdx < queryDescriptors.rows; qIdx++ )
{
if( matcher.isMaskedOut( masks, qIdx ) )
{
if( !compactResult ) // push empty vector
matches.push_back( vector<DMatch>() );
}
else
{
// 1. compute distances between i-th query descriptor and all train descriptors
for( size_t iIdx = 0; iIdx < imgCount; iIdx++ )
{
CV_Assert( DataType<ValueType>::type == matcher.trainDescCollection[iIdx].type() || matcher.trainDescCollection[iIdx].empty() );
CV_Assert( queryDescriptors.cols == matcher.trainDescCollection[iIdx].cols ||
matcher.trainDescCollection[iIdx].empty() );
const ValueType* d1 = (const ValueType*)(queryDescriptors.data + queryDescriptors.step*qIdx);
allDists[iIdx].setTo( Scalar::all(std::numeric_limits<DistanceType>::max()) );
for( int tIdx = 0; tIdx < matcher.trainDescCollection[iIdx].rows; tIdx++ )
{
if( masks.empty() || matcher.isPossibleMatch(masks[iIdx], qIdx, tIdx) )
{
const ValueType* d2 = (const ValueType*)(matcher.trainDescCollection[iIdx].data +
matcher.trainDescCollection[iIdx].step*tIdx);
allDists[iIdx].at<DistanceType>(0, tIdx) = matcher.distance(d1, d2, dimension);
}
}
}
// 2. choose k nearest matches for query[i]
matches.push_back( vector<DMatch>() );
vector<vector<DMatch> >::reverse_iterator curMatches = matches.rbegin();
for( int k = 0; k < knn; k++ )
{
DMatch bestMatch;
bestMatch.distance = std::numeric_limits<float>::max();
for( size_t iIdx = 0; iIdx < imgCount; iIdx++ )
{
if( !allDists[iIdx].empty() )
{
double minVal;
Point minLoc;
minMaxLoc( allDists[iIdx], &minVal, 0, &minLoc, 0 );
if( minVal < bestMatch.distance )
bestMatch = DMatch( qIdx, minLoc.x, (int)iIdx, (float)minVal );
}
}
if( bestMatch.trainIdx == -1 )
break;
allDists[bestMatch.imgIdx].at<DistanceType>(0, bestMatch.trainIdx) = std::numeric_limits<DistanceType>::max();
curMatches->push_back( bestMatch );
}
//TODO should already be sorted at this point?
std::sort( curMatches->begin(), curMatches->end() );
}
}
for( int qIdx = 0; qIdx < queryDescriptors.rows; qIdx++ )
{
if( matcher.isMaskedOut( masks, qIdx ) )
{
if( !compactResult ) // push empty vector
matches.push_back( vector<DMatch>() );
}
else
{
// 1. compute distances between i-th query descriptor and all train descriptors
for( size_t iIdx = 0; iIdx < imgCount; iIdx++ )
{
CV_Assert( DataType<ValueType>::type == matcher.trainDescCollection[iIdx].type() || matcher.trainDescCollection[iIdx].empty() );
CV_Assert( queryDescriptors.cols == matcher.trainDescCollection[iIdx].cols ||
matcher.trainDescCollection[iIdx].empty() );
const ValueType* d1 = (const ValueType*)(queryDescriptors.data + queryDescriptors.step*qIdx);
allDists[iIdx].setTo( Scalar::all(std::numeric_limits<DistanceType>::max()) );
for( int tIdx = 0; tIdx < matcher.trainDescCollection[iIdx].rows; tIdx++ )
{
if( masks.empty() || matcher.isPossibleMatch(masks[iIdx], qIdx, tIdx) )
{
const ValueType* d2 = (const ValueType*)(matcher.trainDescCollection[iIdx].data +
matcher.trainDescCollection[iIdx].step*tIdx);
allDists[iIdx].at<DistanceType>(0, tIdx) = matcher.distance(d1, d2, dimension);
}
}
}
// 2. choose k nearest matches for query[i]
matches.push_back( vector<DMatch>() );
vector<vector<DMatch> >::reverse_iterator curMatches = matches.rbegin();
for( int k = 0; k < knn; k++ )
{
DMatch bestMatch;
bestMatch.distance = std::numeric_limits<float>::max();
for( size_t iIdx = 0; iIdx < imgCount; iIdx++ )
{
if( !allDists[iIdx].empty() )
{
double minVal;
Point minLoc;
minMaxLoc( allDists[iIdx], &minVal, 0, &minLoc, 0 );
if( minVal < bestMatch.distance )
bestMatch = DMatch( qIdx, minLoc.x, (int)iIdx, (float)minVal );
}
}
if( bestMatch.trainIdx == -1 )
break;
allDists[bestMatch.imgIdx].at<DistanceType>(0, bestMatch.trainIdx) = std::numeric_limits<DistanceType>::max();
curMatches->push_back( bestMatch );
}
//TODO should already be sorted at this point?
std::sort( curMatches->begin(), curMatches->end() );
}
}
}
template<class Distance>
......
modules/features2d/src/matchers.cpp
View file @ fcd999ae
......@@ -328,6 +328,10 @@ Ptr<DescriptorMatcher> DescriptorMatcher::create( const string& descriptorMatche
{
dm = new BruteForceMatcher<L2<float> >();
}
else if( !descriptorMatcherType.compare( "BruteForce-SL2" ) ) // Squared L2
{
dm = new BruteForceMatcher<SL2<float> >();
}
else if( !descriptorMatcherType.compare( "BruteForce-L1" ) )
{
dm = new BruteForceMatcher<L1<float> >();
......@@ -345,10 +349,10 @@ Ptr<DescriptorMatcher> DescriptorMatcher::create( const string& descriptorMatche
}
/*
* BruteForce L2 specialization
* BruteForce SL2 and L2 specialization
*/
template<>
void BruteForceMatcher<L2<float> >::knnMatchImpl( const Mat& queryDescriptors, vector<vector<DMatch> >& matches, int knn,
void BruteForceMatcher<SL2<float> >::knnMatchImpl( const Mat& queryDescriptors, vector<vector<DMatch> >& matches, int knn,
const vector<Mat>& masks, bool compactResult )
{
#ifndef HAVE_EIGEN
......@@ -427,7 +431,7 @@ void BruteForceMatcher<L2<float> >::knnMatchImpl( const Mat& queryDescriptors, v
break;
e_allDists[bestImgIdx](bestTrainIdx) = -std::numeric_limits<float>::max();
curMatches->push_back( DMatch(qIdx, bestTrainIdx, bestImgIdx, sqrt((-2)*totalMaxCoeff + queryNorm2)) );
curMatches->push_back( DMatch(qIdx, bestTrainIdx, bestImgIdx, (-2)*totalMaxCoeff + queryNorm2) );
}
std::sort( curMatches->begin(), curMatches->end() );
}
......@@ -436,7 +440,7 @@ void BruteForceMatcher<L2<float> >::knnMatchImpl( const Mat& queryDescriptors, v
}
template<>
void BruteForceMatcher<L2<float> >::radiusMatchImpl( const Mat& queryDescriptors, vector<vector<DMatch> >& matches, float maxDistance,
void BruteForceMatcher<SL2<float> >::radiusMatchImpl( const Mat& queryDescriptors, vector<vector<DMatch> >& matches, float maxDistance,
const vector<Mat>& masks, bool compactResult )
{
#ifndef HAVE_EIGEN
......@@ -492,7 +496,7 @@ void BruteForceMatcher<L2<float> >::radiusMatchImpl( const Mat& queryDescriptors
{
if( masks.empty() || isPossibleMatch(masks[iIdx], qIdx, tIdx) )
{
float d = sqrt((-2)*e_allDists[iIdx](tIdx) + queryNorm2);
float d = (-2)*e_allDists[iIdx](tIdx) + queryNorm2;
if( d < maxDistance )
curMatches->push_back( DMatch( qIdx, tIdx, iIdx, d ) );
}
......@@ -504,6 +508,40 @@ void BruteForceMatcher<L2<float> >::radiusMatchImpl( const Mat& queryDescriptors
#endif
}
inline void sqrtDistance( vector<vector<DMatch> >& matches )
{
for( size_t imgIdx = 0; imgIdx < matches.size(); imgIdx++ )
{
for( size_t matchIdx = 0; matchIdx < matches[imgIdx].size(); matchIdx++ )
{
matches[imgIdx][matchIdx].distance = std::sqrt( matches[imgIdx][matchIdx].distance );
}
}
}
template<>
void BruteForceMatcher<L2<float> >::knnMatchImpl( const Mat& queryDescriptors, vector<vector<DMatch> >& matches, int knn,
const vector<Mat>& masks, bool compactResult )
{
BruteForceMatcher<SL2<float> > matcherSL2;
matcherSL2.add( getTrainDescriptors() );
matcherSL2.knnMatch( queryDescriptors, matches, knn, masks, compactResult );
sqrtDistance( matches );
}
template<>
void BruteForceMatcher<L2<float> >::radiusMatchImpl( const Mat& queryDescriptors, vector<vector<DMatch> >& matches, float maxDistance,
const vector<Mat>& masks, bool compactResult )
{
const float maxDistance2 = maxDistance * maxDistance;
BruteForceMatcher<SL2<float> > matcherSL2;
matcherSL2.add( getTrainDescriptors() );
matcherSL2.radiusMatch( queryDescriptors, matches, maxDistance2, masks, compactResult );
sqrtDistance( matches );
}
/*
* Flann based matcher
*/
......
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