Computer Vision Research

http://cuby5566.pixnet.net/blog/post/44131382-%5B-opencv-%5D-affine-%26-perspective

IplImage* eig_image = cvCreateImage( cvGetSize( iplT0gray ) , IPL_DEPTH_32F , 1 );
IplImage* temp_image = cvCreateImage( cvGetSize( iplT0gray ) , IPL_DEPTH_32F , 1 );
//---------------------------------------------------------
const int MAX_CORNERS = INT_POINT_SIZE;
//---------------------------------------------------------
CvPoint2D32f corners[MAX_CORNERS] = {0}, corners2[MAX_CORNERS] = {0};
//---------------------------------------------------------
double quality_level = 0.05, min_distance = 2, k = 0.04;
int eig_block_size = 3, use_harris = false, corner_count = MAX_CORNERS;
//---------------------------------------------------------
IplImage* pyramid1 = cvCreateImage( cvGetSize( iplT0gray ) , IPL_DEPTH_8U , 1 );
IplImage* pyramid2 = cvCreateImage( cvGetSize( iplT0gray ) , IPL_DEPTH_8U , 1 );
//---------------------------------------------------------
char OF_FoundFeature[MAX_CORNERS];	
float OF_FeatureError[MAX_CORNERS];
CvTermCriteria OF_TerminationCriteria = cvTermCriteria( CV_TERMCRIT_ITER | CV_TERMCRIT_EPS, 20, .3 );
//---------------------------------------------------------
cvGoodFeaturesToTrack( iplT0gray , eig_image, temp_image , corners , & corner_count , quality_level , min_distance , NULL , eig_block_size , use_harris , k );
cvCalcOpticalFlowPyrLK( iplT0gray , iplT1gray , pyramid1 , pyramid2 , corners , corners2 ,
					corner_count , cvSize(3,3) , 5 , OF_FoundFeature , OF_FeatureError, OF_TerminationCriteria , 0 );

Using the corners, we can get these two frame's transformation matrix, and warping back.
* You need to using SIFT or other methods to get these two frame's corners ( feature points ), when these two frame with big change, like scaling, rotation, etc.

1. )  Affine Transformation :
 
CvMat* H = cvCreateMat( 2 , 3 , CV_32FC1 );
cvGetAffineTransform( corners2 , corners , H );
cvWarpAffine( iplSrc , iplObj , H );
 
 
2. )  Perspective Transformation :
 
CvMat* H = cvCreateMat( 3 , 3 , CV_32FC1 );
cvGetPerspectiveTransform( corners2 , corners , H );
cvWarpPerspective( iplSrc , iplObj , H , CV_INTER_LINEAR+CV_WARP_FILL_OUTLIERS , CV_RGB(255,0,0) );
 
 
3. )  Perspective Transformation ( with RANSAC ) :
 
CvMat* H = cvCreateMat( 3 , 3 , CV_32FC1 );
CvMat c = cvMat( 1 , 100 , CV_32FC2 , corners );
CvMat c2 = cvMat( 1 , 100 , CV_32FC2 , corners2 );
cvFindHomography( & c2 , & c , H , CV_RANSAC , 30 );
cvWarpPerspective( iplSrc , iplObj , H , CV_INTER_LINEAR+CV_WARP_FILL_OUTLIERS , CV_RGB(255,0,0) );
 
 
 

Mastering OpenCV with Practical Computer Vision Projects의 3장에 보면 Homography refinement의 예제가 있다. 여러방법이 있겠지만, 책의 예를 보면, 대충 감이 잡히지 않을까?

bool PatternDetector::findPattern(const cv::Mat& image, PatternTrackingInfo& info)

{

    // Convert input image to gray

    getGray(image, m_grayImg);

    // Extract feature points from input gray image

    extractFeatures(m_grayImg, m_queryKeypoints, m_queryDescriptors);

    // Get matches with current pattern

    getMatches(m_queryDescriptors, m_matches);

    // Find homography transformation and detect good matches

    bool homographyFound = refineMatchesWithHomography(

        m_queryKeypoints, 

        m_pattern.keypoints, 

        homographyReprojectionThreshold, 

        m_matches,

        m_roughHomography); 

    if (homographyFound) // 현재 매칭된 Homography을 찾았다.

    {

        // If homography refinement enabled improve found transformation

        if (enableHomographyRefinement)

        {

            // Warp image using found homography

            // m_roughHomography을 이용하여 Warping - Reference Image(완전 사각형) 모양으로 와핑 

            cv::warpPerspective(m_grayImg, m_warpedImg, m_roughHomography, m_pattern.size,  cv::WARP_INVERSE_MAP | cv::INTER_CUBIC);

            // Get refined matches:

            std::vector<cv::KeyPoint> warpedKeypoints;

            std::vector<cv::DMatch> refinedMatches;

// 위에서 와핑한 이미지에 대해 다시 추출하고 매칭

            // Detect features on warped image

            extractFeatures(m_warpedImg, warpedKeypoints, m_queryDescriptors);

            // Match with pattern

            getMatches(m_queryDescriptors, refinedMatches);

            // Estimate new refinement homography

            homographyFound = refineMatchesWithHomography(

                warpedKeypoints, 

                m_pattern.keypoints, 

                homographyReprojectionThreshold, 

                refinedMatches, 

                m_refinedHomography); //  교정된 Homography을 계산

            // Get a result homography as result of matrix product of refined and rough homographies: 

         // 이부분이 조금 이해 안감? 교정된 Homography를 얻으려고 하는 것 같은데? 왜 곱하지?

            info.homography = m_roughHomography * m_refinedHomography;

             // Transform contour with precise homography

           cv::perspectiveTransform(m_pattern.points2d, info.points2d, info.homography);

        }

        else

        { 

            // 그냥 새로운것으로 대체

            info.homography = m_roughHomography;

            // Transform contour with rough homography

            cv::perspectiveTransform(m_pattern.points2d, info.points2d, m_roughHomography);

        }

    }

    return homographyFound;

}