implemented mean shift segmentation with elimination of small segments, added tests

pull/13383/head
Alexey Spizhevoy 14 years ago
parent 5386a8e7cb
commit 970dd7f593
  1. 4
      modules/gpu/include/opencv2/gpu/gpu.hpp
  2. 475
      modules/gpu/src/mssegmentation.cpp
  3. 103
      tests/gpu/src/mssegmentation.cpp

@ -473,6 +473,10 @@ namespace cv
CV_EXPORTS void meanShiftProc(const GpuMat& src, GpuMat& dstr, GpuMat& dstsp, int sp, int sr,
TermCriteria criteria = TermCriteria(TermCriteria::MAX_ITER + TermCriteria::EPS, 5, 1));
//! Does mean shift segmentation with elimiation of small regions.
CV_EXPORTS void meanShiftSegmentation(const GpuMat& src, Mat& dst, int sp, int sr, int minsize,
TermCriteria criteria = TermCriteria(TermCriteria::MAX_ITER + TermCriteria::EPS, 5, 1));
//! Does coloring of disparity image: [0..ndisp) -> [0..240, 1, 1] in HSV.
//! Supported types of input disparity: CV_8U, CV_16S.
//! Output disparity has CV_8UC4 type in BGRA format (alpha = 255).

@ -0,0 +1,475 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include <time.h>
#include <vector>
#include "precomp.hpp"
#if !defined(HAVE_CUDA)
namespace cv
{
namespace gpu
{
void meanShiftSegmentation(const GpuMat&, Mat&, int, int, int, TermCriteria) { throw_nogpu(); }
} // namespace gpu
} // namespace cv
#else
//#define _MSSEGMENTATION_DBG
#ifdef _MSSEGMENTATION_DBG
#include <iostream>
#define LOG(s) std::cout << (s) << std::endl
#define LOG2(s1, s2) std::cout << (s1) << (s2) << std::endl
#define DBG(code) code
#else
#define LOG(s1)
#define LOG2(s1, s2)
#define DBG(code)
#endif
#define PIX(y, x) ((y) * ncols + (x))
using namespace std;
// Auxiliray stuff
namespace
{
//
// Declarations
//
class DjSets
{
public:
DjSets(int n);
~DjSets();
int find(int elem) const;
int merge(int set1, int set2);
int* parent;
int* rank;
int* size;
private:
DjSets(const DjSets&) {}
DjSets operator =(const DjSets&) {}
};
template <typename T>
struct GraphEdge
{
GraphEdge() {}
GraphEdge(int to, int next, const T& val) : to(to), next(next), val(val) {}
int to;
int next;
T val;
};
template <typename T>
class Graph
{
public:
typedef GraphEdge<T> Edge;
Graph(int numv, int nume_max);
~Graph();
void addEdge(int from, int to, const T& val=T());
int* start;
Edge* edges;
int numv;
int nume_max;
int nume;
private:
Graph(const Graph&) {}
Graph operator =(const Graph&) {}
};
struct SegmLinkVal
{
SegmLinkVal() {}
SegmLinkVal(int dr, int dsp) : dr(dr), dsp(dsp) {}
bool operator <(const SegmLinkVal& other) const
{
return dr + dsp < other.dr + other.dsp;
}
int dr;
int dsp;
};
struct SegmLink
{
SegmLink() {}
SegmLink(int from, int to, const SegmLinkVal& val)
: from(from), to(to), val(val) {}
int from;
int to;
SegmLinkVal val;
};
struct SegmLinkCmp
{
bool operator ()(const SegmLink& lhs, const SegmLink& rhs) const
{
return lhs.val < rhs.val;
}
};
//
// Implementation
//
DjSets::DjSets(int n)
{
parent = new int[n];
rank = new int[n];
size = new int[n];
for (int i = 0; i < n; ++i)
{
parent[i] = i;
rank[i] = 0;
size[i] = 1;
}
}
DjSets::~DjSets()
{
delete[] parent;
delete[] rank;
delete[] size;
}
inline int DjSets::find(int elem) const
{
int set = elem;
while (set != parent[set])
set = parent[set];
while (elem != parent[elem])
{
int next = parent[elem];
parent[elem] = set;
elem = next;
}
return set;
}
inline int DjSets::merge(int set1, int set2)
{
if (rank[set1] < rank[set2])
{
parent[set1] = set2;
size[set2] += size[set1];
return set2;
}
if (rank[set2] < rank[set1])
{
parent[set2] = set1;
size[set1] += size[set2];
return set1;
}
parent[set1] = set2;
rank[set2]++;
size[set2] += size[set1];
return set2;
}
template <typename T>
Graph<T>::Graph(int numv, int nume_max)
{
this->numv = numv;
this->nume_max = nume_max;
start = new int[numv];
for (int i = 0; i < numv; ++i)
start[i] = -1;
edges = new Edge[nume_max];
nume = 0;
}
template <typename T>
Graph<T>::~Graph()
{
delete[] start;
delete[] edges;
}
template <typename T>
inline void Graph<T>::addEdge(int from, int to, const T& val)
{
Edge* edge = edges + nume;
new (edge) SegmLink(to, start[from], val);
start[from] = nume;
nume++;
}
inline int sqr(int x)
{
return x * x;
}
} // anonymous namespace
namespace cv
{
namespace gpu
{
void meanShiftSegmentation(const GpuMat& src, Mat& dst, int sp, int sr, int minsize, TermCriteria criteria)
{
CV_Assert(src.type() == CV_8UC4);
const int nrows = src.rows;
const int ncols = src.cols;
const int hr = sr;
const int hsp = sp;
DBG(clock_t start = clock());
// Perform mean shift procedure and obtain region and spatial maps
GpuMat h_rmap, h_spmap;
meanShiftProc(src, h_rmap, h_spmap, sp, sr, criteria);
Mat rmap = h_rmap;
Mat spmap = h_spmap;
LOG2("meanshift:", clock() - start);
DBG(start = clock());
Graph<SegmLinkVal> g(nrows * ncols, 4 * (nrows - 1) * (ncols - 1)
+ (nrows - 1) + (ncols - 1));
LOG2("ragalloc:", clock() - start);
DBG(start = clock());
// Make region adjacent graph from image
// TODO: SSE?
Vec4b r1;
Vec4b r2[4];
Point_<short> sp1;
Point_<short> sp2[4];
int dr[4];
int dsp[4];
for (int y = 0; y < nrows - 1; ++y)
{
Vec4b* ry = rmap.ptr<Vec4b>(y);
Vec4b* ryp = rmap.ptr<Vec4b>(y + 1);
Point_<short>* spy = spmap.ptr<Point_<short> >(y);
Point_<short>* spyp = spmap.ptr<Point_<short> >(y + 1);
for (int x = 0; x < ncols - 1; ++x)
{
r1 = ry[x];
sp1 = spy[x];
r2[0] = ry[x + 1];
r2[1] = ryp[x];
r2[2] = ryp[x + 1];
r2[3] = ryp[x];
sp2[0] = spy[x + 1];
sp2[1] = spyp[x];
sp2[2] = spyp[x + 1];
sp2[3] = spyp[x];
dr[0] = sqr(r1[0] - r2[0][0]) + sqr(r1[1] - r2[0][1]) + sqr(r1[2] - r2[0][2]);
dr[1] = sqr(r1[0] - r2[1][0]) + sqr(r1[1] - r2[1][1]) + sqr(r1[2] - r2[1][2]);
dr[2] = sqr(r1[0] - r2[2][0]) + sqr(r1[1] - r2[2][1]) + sqr(r1[2] - r2[2][2]);
dsp[0] = sqr(sp1.x - sp2[0].x) + sqr(sp1.y - sp2[0].y);
dsp[1] = sqr(sp1.x - sp2[1].x) + sqr(sp1.y - sp2[1].y);
dsp[2] = sqr(sp1.x - sp2[2].x) + sqr(sp1.y - sp2[2].y);
r1 = ry[x + 1];
sp1 = spy[x + 1];
dr[3] = sqr(r1[0] - r2[3][0]) + sqr(r1[1] - r2[3][1]) + sqr(r1[2] - r2[3][2]);
dsp[3] = sqr(sp1.x - sp2[3].x) + sqr(sp1.y - sp2[3].y);
g.addEdge(PIX(y, x), PIX(y, x + 1), SegmLinkVal(dr[0], dsp[0]));
g.addEdge(PIX(y, x), PIX(y + 1, x), SegmLinkVal(dr[1], dsp[1]));
g.addEdge(PIX(y, x), PIX(y + 1, x + 1), SegmLinkVal(dr[2], dsp[2]));
g.addEdge(PIX(y, x + 1), PIX(y, x + 1), SegmLinkVal(dr[3], dsp[3]));
}
}
for (int y = 0; y < nrows - 1; ++y)
{
r1 = rmap.at<Vec4b>(y, ncols - 1);
r2[0] = rmap.at<Vec4b>(y + 1, ncols - 1);
sp1 = spmap.at<Point_<short> >(y, ncols - 1);
sp2[0] = spmap.at<Point_<short> >(y + 1, ncols - 1);
dr[0] = sqr(r1[0] - r2[0][0]) + sqr(r1[1] - r2[0][1]) + sqr(r1[2] - r2[0][2]);
dsp[0] = sqr(sp1.x - sp2[0].x) + sqr(sp1.y - sp2[0].y);
g.addEdge(PIX(y, ncols - 1), PIX(y + 1, ncols - 1), SegmLinkVal(dr[0], dsp[0]));
}
for (int x = 0; x < ncols - 1; ++x)
{
r1 = rmap.at<Vec4b>(nrows - 1, x);
r2[0] = rmap.at<Vec4b>(nrows - 1, x + 1);
sp1 = spmap.at<Point_<short> >(nrows - 1, x);
sp2[0] = spmap.at<Point_<short> >(nrows - 1, x + 1);
dr[0] = sqr(r1[0] - r2[0][0]) + sqr(r1[1] - r2[0][1]) + sqr(r1[2] - r2[0][2]);
dsp[0] = sqr(sp1.x - sp2[0].x) + sqr(sp1.y - sp2[0].y);
g.addEdge(PIX(nrows - 1, x), PIX(nrows - 1, x + 1), SegmLinkVal(dr[0], dsp[0]));
}
LOG2("raginit:", clock() - start);
DBG(start = clock());
DjSets comps(g.numv);
LOG2("djsetinit:", clock() - start);
DBG(start = clock());
// Find adjacent components
for (int v = 0; v < g.numv; ++v)
{
for (int e_it = g.start[v]; e_it != -1; e_it = g.edges[e_it].next)
{
int comp1 = comps.find(v);
int comp2 = comps.find(g.edges[e_it].to);
if (comp1 != comp2 && g.edges[e_it].val.dr < hr
&& g.edges[e_it].val.dsp < hsp)
comps.merge(comp1, comp2);
}
}
LOG2("findadjacent:", clock() - start);
DBG(start = clock());
vector<SegmLink> edges;
edges.reserve(g.numv);
LOG2("initedges:", clock() - start);
DBG(start = clock());
for (int v = 0; v < g.numv; ++v)
{
int comp1 = comps.find(v);
for (int e_it = g.start[v]; e_it != -1; e_it = g.edges[e_it].next)
{
int comp2 = comps.find(g.edges[e_it].to);
if (comp1 != comp2)
edges.push_back(SegmLink(comp1, comp2, g.edges[e_it].val));
}
}
LOG2("prepareforsort:", clock() - start);
DBG(start = clock());
// Sort all graph's edges connecting differnet components (in asceding order)
sort(edges.begin(), edges.end(), SegmLinkCmp());
LOG2("sortedges:", clock() - start);
DBG(start = clock());
// Exclude small components (starting from the nearest couple)
vector<SegmLink>::iterator e_it = edges.begin();
for (; e_it != edges.end(); ++e_it)
{
int comp1 = comps.find(e_it->from);
int comp2 = comps.find(e_it->to);
if (comp1 != comp2 && (comps.size[comp1] < minsize || comps.size[comp2] < minsize))
comps.merge(comp1, comp2);
}
LOG2("excludesmall:", clock() - start);
DBG(start = clock());
// Compute sum of the pixel's colors which are in the same segment
Mat h_src = src;
vector<Vec4i> sumcols(nrows * ncols, Vec4i(0, 0, 0, 0));
for (int y = 0; y < nrows; ++y)
{
Vec4b* h_srcy = h_src.ptr<Vec4b>(y);
for (int x = 0; x < ncols; ++x)
{
int parent = comps.find(PIX(y, x));
Vec4b col = h_srcy[x];
Vec4i& sumcol = sumcols[parent];
sumcol[0] += col[0];
sumcol[1] += col[1];
sumcol[2] += col[2];
}
}
LOG2("computesum:", clock() - start);
DBG(start = clock());
// Create final image, color of each segment is the average color of its pixels
dst.create(src.size(), src.type());
for (int y = 0; y < nrows; ++y)
{
Vec4b* dsty = dst.ptr<Vec4b>(y);
for (int x = 0; x < ncols; ++x)
{
int parent = comps.find(PIX(y, x));
const Vec4i& sumcol = sumcols[parent];
Vec4b& dstcol = dsty[x];
dstcol[0] = static_cast<uchar>(sumcol[0] / comps.size[parent]);
dstcol[1] = static_cast<uchar>(sumcol[1] / comps.size[parent]);
dstcol[2] = static_cast<uchar>(sumcol[2] / comps.size[parent]);
}
}
LOG2("createfinal:", clock() - start);
}
} // namespace gpu
} // namespace cv
#endif // #if !defined (HAVE_CUDA)

@ -0,0 +1,103 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// Intel License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000, Intel Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of Intel Corporation may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include <opencv2/opencv.hpp>
#include <opencv2/gpu/gpu.hpp>
#include <iostream>
#include <string>
#include <iosfwd>
#include "gputest.hpp"
using namespace cv;
using namespace cv::gpu;
using namespace std;
struct CV_GpuMeanShiftSegmentationTest : public CvTest {
CV_GpuMeanShiftSegmentationTest() : CvTest( "GPU-MeanShiftSegmentation", "MeanShiftSegmentation" ) {}
void run(int)
{
try
{
Mat img_rgb = imread(string(ts->get_data_path()) + "meanshift/cones.png");
if (img_rgb.empty())
{
ts->set_failed_test_info(CvTS::FAIL_MISSING_TEST_DATA);
return;
}
Mat img;
cvtColor(img_rgb, img, CV_BGR2BGRA);
for (int minsize = 0; minsize < 2000; minsize = (minsize + 1) * 2)
{
stringstream path;
path << ts->get_data_path() << "meanshift/cones_segmented_sp10_sr10_minsize" << minsize << ".png";
Mat dst;
meanShiftSegmentation((GpuMat)img, dst, 10, 10, minsize);
Mat dst_rgb;
cvtColor(dst, dst_rgb, CV_BGRA2BGR);
//imwrite(path.str(), dst_rgb);
Mat dst_ref = imread(path.str());
if (dst_ref.empty())
{
ts->set_failed_test_info(CvTS::FAIL_MISSING_TEST_DATA);
return;
}
if (abs(cv::norm(dst_rgb - dst_ref, NORM_INF)) > 1e-3)
{
ts->printf(CvTS::LOG, "\ndiffers from image *minsize%d.png\n", minsize);
ts->set_failed_test_info(CvTS::FAIL_BAD_ACCURACY);
return;
}
}
}
catch (const cv::Exception& e)
{
if (!check_and_treat_gpu_exception(e, ts))
throw;
return;
}
ts->set_failed_test_info(CvTS::OK);
}
} ms_segm_test;
Loading…
Cancel
Save