使用OpenCV显示SiftGPU提取的特征点和特征匹配

使用OpenCV显示SiftGPU提取的特征点和特征匹配，主要包括以下几个步骤：

1. SiftGPU特征点提取
2. OpenCV读入图像
3. SiftGPU特征点类型到OpenCV特征点类型KeyPoint的转换
4. OpenCV特征点显示
5. SiftGPU描述子类型到OpenCV描述子类型的转换
6. SiftGPU特征匹配
7. SiftGPU特征匹配到OpenCV匹配DMatch类型的转换

具体代码如下:
代码采用image.txt输入待处理的图像路径，此处根据自己实际处理的图像自行设置。

#include 
#include 
#include 
#include 
#include 
#include 
#include   
#include   
#include 

#include 

#include "compare.h"

using namespace std;
using namespace cv;

#if !defined(SIFTGPU_STATIC) && !defined(SIFTGPU_DLL_RUNTIME) 
// SIFTGPU_STATIC comes from compiler
#define SIFTGPU_DLL_RUNTIME
// Load at runtime if the above macro defined
// comment the macro above to use static linking
#endif

#ifdef _WIN32
	#ifdef SIFTGPU_DLL_RUNTIME
		#define WIN32_LEAN_AND_MEAN
		#include 
		#define FREE_MYLIB FreeLibrary
		#define GET_MYPROC GetProcAddress
	#else
		//define this to get dll import definition for win32
		#define SIFTGPU_DLL
		#ifdef _DEBUG 
			#pragma comment(lib, "../lib/SIFTGPU.lib")
		#else
			#pragma comment(lib, "../lib/SIFTGPU.lib")
		#endif
	#endif
#else
	#ifdef SIFTGPU_DLL_RUNTIME
	#include 
	#define FREE_MYLIB dlclose
	#define GET_MYPROC dlsym
	#endif
#endif

#include 

int main(int argc, char** argv)
{
#ifdef SIFTGPU_DLL_RUNTIME
	#ifdef _WIN32
		#ifdef _DEBUG
				HMODULE  hsiftgpu = LoadLibrary("SiftGPU_d.dll");
		#else
				HMODULE  hsiftgpu = LoadLibrary("siftgpu.dll");
		#endif
	#else
			void * hsiftgpu = dlopen("libsiftgpu.so", RTLD_LAZY);
	#endif
			if (hsiftgpu == NULL) {
				DWORD error_id = GetLastError();
				cout << "error loading SIFTGPU.dll" << endl;
				cout << "error id is " << error_id << endl;
				system("pause");
				return 0;
			}
#endif

	//声明SiftGPU并初始化  
	SiftGPU sift;
	char* myargv[4] = { "-fo", "-1", "-v", "1" };
	sift.ParseParam(4, myargv);
	//检查硬件是否支持SiftGPU  
	int support = sift.CreateContextGL();
	if (support != SiftGPU::SIFTGPU_FULL_SUPPORTED)
	{
		cerr << "SiftGPU is not supported!" << endl;
		return 2;
	}
	//读入txt文件列表内的图像名
	vector<string> imageNames;
	string fileName;
	int imgCount = 0;
	ifstream fin("image.txt");
	while (getline(fin, fileName)){
		imgCount++;
		imageNames.push_back(fileName);
	}
	if (imgCount == 0){
		cerr << "cannot get image files" << endl;
		return 1;
	}
	else{
		cout << "Read In " << imgCount << "  Pictures" << endl;
	}
	//OpenCV读入图像
	vector<Mat> images;
	for (int i = 0; i < imgCount; ++i){
		Mat temp = imread(imageNames[i]);
		images.push_back(temp);
	}
	cout << "OpenCV Read in " << images.size() << " images" << endl;
	//SiftGPU读入图像并处理
	vector<int> num;
	vector<vector<float>> descriptors;
	vector<vector<SiftGPU::SiftKeypoint>> keys;
	for (int i = 0; i < imgCount; ++i){
		const char* temp = imageNames[i].data();
		sift.RunSIFT(temp);
		int _num = sift.GetFeatureNum();
		num.push_back(_num);
		cout << "Keypoints Number of " << imageNames[i] << " is= " << _num << endl << endl;
		vector<SiftGPU::SiftKeypoint> _keys;
		vector<float> _descriptors;
		_keys.resize(_num);
		_descriptors.resize(128*_num);
		sift.GetFeatureVector(&_keys[0], &_descriptors[0]);	
		keys.push_back(_keys);
		descriptors.push_back(_descriptors);
	}
	//SiftGPU的特征点赋值于OpenCV关键点类型
	vector<vector<KeyPoint>> cvKeypoints;
	for (int i = 0; i < imgCount; ++i){
		vector<KeyPoint> cvKeys;
		for (int j = 0; j < keys[i].size(); ++j){
			KeyPoint cvTempKey;
			cvTempKey.pt.x = keys[i][j].x;
			cvTempKey.pt.y = keys[i][j].y;
			cvKeys.push_back(cvTempKey);
		}
		cvKeypoints.push_back(cvKeys);
	}
	//验证第i张图片的特征点数是否相同
	for (int i = 0; i < imgCount; ++i){
		if (cvKeypoints[i].size() != keys[i].size())
			cerr << "Image " << i << "'s Keypoints Number isn't correct." << endl;
	}
	//使用OpenCV接口显示关键点
	for (int i = 0; i < imgCount; ++i){
		Mat featureImage;
		drawKeypoints(images[i], cvKeypoints[i], featureImage, Scalar(255, 255, 255), DrawMatchesFlags::DEFAULT);
		namedWindow("Sift Keypoints",WINDOW_NORMAL);
		imshow("Sift Keypoints", featureImage);
		waitKey(0);
	}

	//SiftGPU的描述子赋值于OpenCV描述子类型
	vector<Mat> cvDescriptors(imgCount);  //每个Mat内部是num行，128列的特征点描述矩阵
	for (int inum = 0; inum < imgCount; ++inum){
		//图i的SiftGPU的描述子_des
		vector<vector<float>> _des;
		for (int sj = 0; sj < num[inum]; ++sj){
			vector<float> temp(128, 0.0);
			for (int sk = 0; sk < 128; ++sk){
				temp[sk] = descriptors[inum][sk + 128 * sj];
			}
			_des.push_back(temp);
		}
		//图i的OpenCV描述子_cvDes
		Mat _cvDes(num[inum], 128, CV_32F);		
		for (int cj = 0; cj < num[inum]; ++cj){
			float* pxDesMat = _cvDes.ptr<float>(cj);
			for (int ck = 0; ck < 128; ++ck){
				pxDesMat[ck] = _des[cj][ck];
			}
		}
		输出_des的极值
		//float smax = _des[0][0];
		//float smin = _des[0][0];
		//for (int simm = 0; simm < _des.size(); ++simm){
		//	for (int sjmm = 0; sjmm < 128; ++sjmm){
		//		if (smax < _des[simm][sjmm])
		//			smax = _des[simm][sjmm];
		//		if (smin > _des[simm][sjmm])
		//			smin = _des[simm][sjmm];
		//	}
		//}
		//cout << "SiftGPU Descriptors Max= " << smax << " , Min=" << smin << endl;
		输出_cvDes的极值
		//float cmax = _cvDes.at(0, 0);
		//float cmin = _cvDes.at(0, 0);
		//for (int cimm = 0; cimm < _cvDes.rows; ++cimm){
		//	for (int cjmm = 0; cjmm < 128; ++cjmm){
		//		if (cmax < _cvDes.at(cimm,cjmm))
		//			cmax = _cvDes.at(cimm, cjmm);
		//		if (cmin > _cvDes.at(cimm, cjmm))
		//			cmin = _cvDes.at(cimm, cjmm);
		//	}
		//}
		//cout << "OpenCV Descriptors Max= " << cmax << " , Min=" << cmin << endl;
		cvDescriptors.push_back(_cvDes);
	}
	map<pair<int, int>, vector<DMatch>> matches_matrix;
	map<pair<int, int>, int> matches_num;
	SiftMatchGPU matcher;
	matcher.VerifyContextGL();
	for (int i = 0; i < imgCount - 1; ++i){
		for (int j = i + 1; j < imgCount; ++j){
			matcher.SetDescriptors(0, num[i], &descriptors[i][0]);
			matcher.SetDescriptors(1, num[j], &descriptors[j][0]);
			int(*match_buf)[2] = new int[num[i]][2];
			int num_match = matcher.GetSiftMatch(num[i], match_buf);
			matches_num[make_pair(i, j)] = num_match;
			cout << "Matches between Image" << i << " and Image" << j << " are " << num_match << endl;
			//定义OpenCV的match类型，并赋值quryIdx和trainIdx
			vector<DMatch> matches_tmp;
			for (int mk = 0; mk < num_match; ++mk){
				DMatch match;
				match.queryIdx = match_buf[mk][0];
				match.trainIdx = match_buf[mk][1];
				Point2f p1, p2;
				p1.x = cvKeypoints[i][match.queryIdx].pt.x;
				p1.y = cvKeypoints[i][match.queryIdx].pt.y;
				p2.x = cvKeypoints[j][match.trainIdx].pt.x;
				p2.y = cvKeypoints[j][match.trainIdx].pt.y;
				float dist;
				dist = (float)sqrtf(((p1.x - p2.x)*(p1.x - p2.x) + (p1.y - p2.y)*(p1.y - p2.y)));
				match.distance = dist;
				//cout << "Image" << i << " & Image" << j << " Match " << mk << ": " << endl;
				//cout << "Idx: queryIdx is= " << match.queryIdx << " , trainIdx is= " << match.trainIdx << endl;
				//cout << "                         Keypoint1: " << p1.x << "   " << p1.y << endl;
				//cout << "                         Keypoint2: " << p2.x << "   " << p2.y << endl;
				//cout << "                         Distance: " << match.distance << endl;
				matches_tmp.push_back(match);
			}
			matches_matrix[make_pair(i, j)] = matches_tmp;
			delete[] match_buf;
			//OpenCV匹配显示
			Mat cvImgMatches;
			drawMatches(images[i], cvKeypoints[i], images[j], cvKeypoints[j], matches_tmp, cvImgMatches);
			char showName[100];
			sprintf(showName, "%s%d%s%d", "Matches between ", i, " and ", j);
			namedWindow(showName, WINDOW_NORMAL);
			imshow(showName, cvImgMatches);
			waitKey(0);
		}
	}
	system("pause");
	return 0;
}

使用两幅Kermit图像进行测试的显示结果：
控制台输出：