/* * cvoneway.cpp * one_way_sample * * Created by Victor Eruhimov on 3/23/10. * Copyright 2010 Argus Corp. All rights reserved. * */ #include "precomp.hpp" #include "opencv2/highgui/highgui.hpp" #include namespace cv{ inline int round(float value) { if(value > 0) { return int(value + 0.5f); } else { return int(value - 0.5f); } } inline CvRect resize_rect(CvRect rect, float alpha) { return cvRect(rect.x + round((float)(0.5*(1 - alpha)*rect.width)), rect.y + round((float)(0.5*(1 - alpha)*rect.height)), round(rect.width*alpha), round(rect.height*alpha)); } CvMat* ConvertImageToMatrix(IplImage* patch); class CvCameraPose { public: CvCameraPose() { m_rotation = cvCreateMat(1, 3, CV_32FC1); m_translation = cvCreateMat(1, 3, CV_32FC1); }; ~CvCameraPose() { cvReleaseMat(&m_rotation); cvReleaseMat(&m_translation); }; void SetPose(CvMat* rotation, CvMat* translation) { cvCopy(rotation, m_rotation); cvCopy(translation, m_translation); }; CvMat* GetRotation() {return m_rotation;}; CvMat* GetTranslation() {return m_translation;}; protected: CvMat* m_rotation; CvMat* m_translation; }; // AffineTransformPatch: generates an affine transformed image patch. // - src: source image (roi is supported) // - dst: output image. ROI of dst image should be 2 times smaller than ROI of src. // - pose: parameters of an affine transformation void AffineTransformPatch(IplImage* src, IplImage* dst, CvAffinePose pose); // GenerateAffineTransformFromPose: generates an affine transformation matrix from CvAffinePose instance // - size: the size of image patch // - pose: affine transformation // - transform: 2x3 transformation matrix void GenerateAffineTransformFromPose(CvSize size, CvAffinePose pose, CvMat* transform); // Generates a random affine pose CvAffinePose GenRandomAffinePose(); const static int num_mean_components = 500; const static float noise_intensity = 0.15f; static inline CvPoint rect_center(CvRect rect) { return cvPoint(rect.x + rect.width/2, rect.y + rect.height/2); } void homography_transform(IplImage* frontal, IplImage* result, CvMat* homography) { cvWarpPerspective(frontal, result, homography); } CvAffinePose perturbate_pose(CvAffinePose pose, float noise) { // perturbate the matrix float noise_mult_factor = 1 + (0.5f - float(rand())/RAND_MAX)*noise; float noise_add_factor = noise_mult_factor - 1; CvAffinePose pose_pert = pose; pose_pert.phi += noise_add_factor; pose_pert.theta += noise_mult_factor; pose_pert.lambda1 *= noise_mult_factor; pose_pert.lambda2 *= noise_mult_factor; return pose_pert; } void generate_mean_patch(IplImage* frontal, IplImage* result, CvAffinePose pose, int pose_count, float noise) { IplImage* sum = cvCreateImage(cvSize(result->width, result->height), IPL_DEPTH_32F, 1); IplImage* workspace = cvCloneImage(result); IplImage* workspace_float = cvCloneImage(sum); cvSetZero(sum); for(int i = 0; i < pose_count; i++) { CvAffinePose pose_pert = perturbate_pose(pose, noise); AffineTransformPatch(frontal, workspace, pose_pert); cvConvertScale(workspace, workspace_float); cvAdd(sum, workspace_float, sum); } cvConvertScale(sum, result, 1.0f/pose_count); cvReleaseImage(&workspace); cvReleaseImage(&sum); cvReleaseImage(&workspace_float); } void generate_mean_patch_fast(IplImage* /*frontal*/, IplImage* /*result*/, CvAffinePose /*pose*/, CvMat* /*pca_hr_avg*/, CvMat* /*pca_hr_eigenvectors*/, const OneWayDescriptor* /*pca_descriptors*/) { /*for(int i = 0; i < pca_hr_eigenvectors->cols; i++) { }*/ } void readPCAFeatures(const char *filename, CvMat** avg, CvMat** eigenvectors, const char *postfix = ""); void readPCAFeatures(const FileNode &fn, CvMat** avg, CvMat** eigenvectors, const char* postfix = ""); void savePCAFeatures(FileStorage &fs, const char* postfix, CvMat* avg, CvMat* eigenvectors); void calcPCAFeatures(vector& patches, FileStorage &fs, const char* postfix, CvMat** avg, CvMat** eigenvectors); void loadPCAFeatures(const char* path, const char* images_list, vector& patches, CvSize patch_size); void generatePCAFeatures(const char* path, const char* img_filename, FileStorage& fs, const char* postfix, CvSize patch_size, CvMat** avg, CvMat** eigenvectors); void eigenvector2image(CvMat* eigenvector, IplImage* img); void FindOneWayDescriptor(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, int& desc_idx, int& pose_idx, float& distance, CvMat* avg = 0, CvMat* eigenvalues = 0); void FindOneWayDescriptor(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, int n, std::vector& desc_idxs, std::vector& pose_idxs, std::vector& distances, CvMat* avg = 0, CvMat* eigenvalues = 0); void FindOneWayDescriptor(cv::flann::Index* m_pca_descriptors_tree, CvSize patch_size, int m_pca_dim_low, int m_pose_count, IplImage* patch, int& desc_idx, int& pose_idx, float& distance, CvMat* avg = 0, CvMat* eigenvalues = 0); void FindOneWayDescriptorEx(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, float scale_min, float scale_max, float scale_step, int& desc_idx, int& pose_idx, float& distance, float& scale, CvMat* avg, CvMat* eigenvectors); void FindOneWayDescriptorEx(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, float scale_min, float scale_max, float scale_step, int n, std::vector& desc_idxs, std::vector& pose_idxs, std::vector& distances, std::vector& scales, CvMat* avg, CvMat* eigenvectors); void FindOneWayDescriptorEx(cv::flann::Index* m_pca_descriptors_tree, CvSize patch_size, int m_pca_dim_low, int m_pose_count, IplImage* patch, float scale_min, float scale_max, float scale_step, int& desc_idx, int& pose_idx, float& distance, float& scale, CvMat* avg, CvMat* eigenvectors); inline CvRect fit_rect_roi_fixedsize(CvRect rect, CvRect roi) { CvRect fit = rect; fit.x = MAX(fit.x, roi.x); fit.y = MAX(fit.y, roi.y); fit.x = MIN(fit.x, roi.x + roi.width - fit.width - 1); fit.y = MIN(fit.y, roi.y + roi.height - fit.height - 1); return(fit); } inline CvRect fit_rect_fixedsize(CvRect rect, IplImage* img) { CvRect roi = cvGetImageROI(img); return fit_rect_roi_fixedsize(rect, roi); } OneWayDescriptor::OneWayDescriptor() { m_pose_count = 0; m_samples = 0; m_input_patch = 0; m_train_patch = 0; m_pca_coeffs = 0; m_affine_poses = 0; m_transforms = 0; m_pca_dim_low = 100; m_pca_dim_high = 100; } OneWayDescriptor::~OneWayDescriptor() { if(m_pose_count) { for(int i = 0; i < m_pose_count; i++) { cvReleaseImage(&m_samples[i]); cvReleaseMat(&m_pca_coeffs[i]); } cvReleaseImage(&m_input_patch); cvReleaseImage(&m_train_patch); delete []m_samples; delete []m_pca_coeffs; if(!m_transforms) { delete []m_affine_poses; } } } void OneWayDescriptor::Allocate(int pose_count, CvSize size, int nChannels) { m_pose_count = pose_count; m_samples = new IplImage* [m_pose_count]; m_pca_coeffs = new CvMat* [m_pose_count]; m_patch_size = cvSize(size.width/2, size.height/2); if(!m_transforms) { m_affine_poses = new CvAffinePose[m_pose_count]; } int length = m_pca_dim_low;//roi.width*roi.height; for(int i = 0; i < m_pose_count; i++) { m_samples[i] = cvCreateImage(cvSize(size.width/2, size.height/2), IPL_DEPTH_32F, nChannels); m_pca_coeffs[i] = cvCreateMat(1, length, CV_32FC1); } m_input_patch = cvCreateImage(GetPatchSize(), IPL_DEPTH_8U, 1); m_train_patch = cvCreateImage(GetInputPatchSize(), IPL_DEPTH_8U, 1); } void cvmSet2DPoint(CvMat* matrix, int row, int col, CvPoint2D32f point) { cvmSet(matrix, row, col, point.x); cvmSet(matrix, row, col + 1, point.y); } void cvmSet3DPoint(CvMat* matrix, int row, int col, CvPoint3D32f point) { cvmSet(matrix, row, col, point.x); cvmSet(matrix, row, col + 1, point.y); cvmSet(matrix, row, col + 2, point.z); } CvAffinePose GenRandomAffinePose() { const float scale_min = 0.8f; const float scale_max = 1.2f; CvAffinePose pose; pose.theta = float(rand())/RAND_MAX*120 - 60; pose.phi = float(rand())/RAND_MAX*360; pose.lambda1 = scale_min + float(rand())/RAND_MAX*(scale_max - scale_min); pose.lambda2 = scale_min + float(rand())/RAND_MAX*(scale_max - scale_min); return pose; } void GenerateAffineTransformFromPose(CvSize size, CvAffinePose pose, CvMat* transform) { CvMat* temp = cvCreateMat(3, 3, CV_32FC1); CvMat* final = cvCreateMat(3, 3, CV_32FC1); cvmSet(temp, 2, 0, 0.0f); cvmSet(temp, 2, 1, 0.0f); cvmSet(temp, 2, 2, 1.0f); CvMat rotation; cvGetSubRect(temp, &rotation, cvRect(0, 0, 3, 2)); cv2DRotationMatrix(cvPoint2D32f(size.width/2, size.height/2), pose.phi, 1.0, &rotation); cvCopy(temp, final); cvmSet(temp, 0, 0, pose.lambda1); cvmSet(temp, 0, 1, 0.0f); cvmSet(temp, 1, 0, 0.0f); cvmSet(temp, 1, 1, pose.lambda2); cvmSet(temp, 0, 2, size.width/2*(1 - pose.lambda1)); cvmSet(temp, 1, 2, size.height/2*(1 - pose.lambda2)); cvMatMul(temp, final, final); cv2DRotationMatrix(cvPoint2D32f(size.width/2, size.height/2), pose.theta - pose.phi, 1.0, &rotation); cvMatMul(temp, final, final); cvGetSubRect(final, &rotation, cvRect(0, 0, 3, 2)); cvCopy(&rotation, transform); cvReleaseMat(&temp); cvReleaseMat(&final); } void AffineTransformPatch(IplImage* src, IplImage* dst, CvAffinePose pose) { CvRect src_large_roi = cvGetImageROI(src); IplImage* temp = cvCreateImage(cvSize(src_large_roi.width, src_large_roi.height), IPL_DEPTH_32F, src->nChannels); cvSetZero(temp); IplImage* temp2 = cvCloneImage(temp); CvMat* rotation_phi = cvCreateMat(2, 3, CV_32FC1); CvSize new_size = cvSize(cvRound(temp->width*pose.lambda1), cvRound(temp->height*pose.lambda2)); IplImage* temp3 = cvCreateImage(new_size, IPL_DEPTH_32F, src->nChannels); cvConvertScale(src, temp); cvResetImageROI(temp); cv2DRotationMatrix(cvPoint2D32f(temp->width/2, temp->height/2), pose.phi, 1.0, rotation_phi); cvWarpAffine(temp, temp2, rotation_phi); cvSetZero(temp); cvResize(temp2, temp3); cv2DRotationMatrix(cvPoint2D32f(temp3->width/2, temp3->height/2), pose.theta - pose.phi, 1.0, rotation_phi); cvWarpAffine(temp3, temp, rotation_phi); cvSetImageROI(temp, cvRect(temp->width/2 - src_large_roi.width/4, temp->height/2 - src_large_roi.height/4, src_large_roi.width/2, src_large_roi.height/2)); cvConvertScale(temp, dst); cvReleaseMat(&rotation_phi); cvReleaseImage(&temp3); cvReleaseImage(&temp2); cvReleaseImage(&temp); } void OneWayDescriptor::GenerateSamples(int pose_count, IplImage* frontal, int norm) { /* if(m_transforms) { GenerateSamplesWithTransforms(pose_count, frontal); return; } */ CvRect roi = cvGetImageROI(frontal); IplImage* patch_8u = cvCreateImage(cvSize(roi.width/2, roi.height/2), frontal->depth, frontal->nChannels); for(int i = 0; i < pose_count; i++) { if(!m_transforms) { m_affine_poses[i] = GenRandomAffinePose(); } //AffineTransformPatch(frontal, patch_8u, m_affine_poses[i]); generate_mean_patch(frontal, patch_8u, m_affine_poses[i], num_mean_components, noise_intensity); double scale = 1.0f; if(norm) { double sum = cvSum(patch_8u).val[0]; scale = 1/sum; } cvConvertScale(patch_8u, m_samples[i], scale); #if 0 double maxval; cvMinMaxLoc(m_samples[i], 0, &maxval); IplImage* test = cvCreateImage(cvSize(roi.width/2, roi.height/2), IPL_DEPTH_8U, 1); cvConvertScale(m_samples[i], test, 255.0/maxval); cvNamedWindow("1", 1); cvShowImage("1", test); cvWaitKey(0); #endif } cvReleaseImage(&patch_8u); } void OneWayDescriptor::GenerateSamplesFast(IplImage* frontal, CvMat* pca_hr_avg, CvMat* pca_hr_eigenvectors, OneWayDescriptor* pca_descriptors) { CvRect roi = cvGetImageROI(frontal); if(roi.width != GetInputPatchSize().width || roi.height != GetInputPatchSize().height) { cvResize(frontal, m_train_patch); frontal = m_train_patch; } CvMat* pca_coeffs = cvCreateMat(1, pca_hr_eigenvectors->cols, CV_32FC1); double maxval; cvMinMaxLoc(frontal, 0, &maxval); CvMat* frontal_data = ConvertImageToMatrix(frontal); double sum = cvSum(frontal_data).val[0]; cvConvertScale(frontal_data, frontal_data, 1.0f/sum); cvProjectPCA(frontal_data, pca_hr_avg, pca_hr_eigenvectors, pca_coeffs); for(int i = 0; i < m_pose_count; i++) { cvSetZero(m_samples[i]); for(int j = 0; j < m_pca_dim_high; j++) { double coeff = cvmGet(pca_coeffs, 0, j); IplImage* patch = pca_descriptors[j + 1].GetPatch(i); cvAddWeighted(m_samples[i], 1.0, patch, coeff, 0, m_samples[i]); #if 0 printf("coeff%d = %f\n", j, coeff); IplImage* test = cvCreateImage(cvSize(12, 12), IPL_DEPTH_8U, 1); double maxval; cvMinMaxLoc(patch, 0, &maxval); cvConvertScale(patch, test, 255.0/maxval); cvNamedWindow("1", 1); cvShowImage("1", test); cvWaitKey(0); #endif } cvAdd(pca_descriptors[0].GetPatch(i), m_samples[i], m_samples[i]); double sum = cvSum(m_samples[i]).val[0]; cvConvertScale(m_samples[i], m_samples[i], 1.0/sum); #if 0 IplImage* test = cvCreateImage(cvSize(12, 12), IPL_DEPTH_8U, 1); /* IplImage* temp1 = cvCreateImage(cvSize(12, 12), IPL_DEPTH_32F, 1); eigenvector2image(pca_hr_avg, temp1); IplImage* test = cvCreateImage(cvSize(12, 12), IPL_DEPTH_8U, 1); cvAdd(m_samples[i], temp1, temp1); cvMinMaxLoc(temp1, 0, &maxval); cvConvertScale(temp1, test, 255.0/maxval);*/ cvMinMaxLoc(m_samples[i], 0, &maxval); cvConvertScale(m_samples[i], test, 255.0/maxval); cvNamedWindow("1", 1); cvShowImage("1", frontal); cvNamedWindow("2", 1); cvShowImage("2", test); cvWaitKey(0); #endif } cvReleaseMat(&pca_coeffs); cvReleaseMat(&frontal_data); } void OneWayDescriptor::SetTransforms(CvAffinePose* poses, CvMat** transforms) { if(m_affine_poses) { delete []m_affine_poses; } m_affine_poses = poses; m_transforms = transforms; } void OneWayDescriptor::Initialize(int pose_count, IplImage* frontal, const char* feature_name, int norm) { m_feature_name = std::string(feature_name); CvRect roi = cvGetImageROI(frontal); m_center = rect_center(roi); Allocate(pose_count, cvSize(roi.width, roi.height), frontal->nChannels); GenerateSamples(pose_count, frontal, norm); } void OneWayDescriptor::InitializeFast(int pose_count, IplImage* frontal, const char* feature_name, CvMat* pca_hr_avg, CvMat* pca_hr_eigenvectors, OneWayDescriptor* pca_descriptors) { if(pca_hr_avg == 0) { Initialize(pose_count, frontal, feature_name, 1); return; } m_feature_name = std::string(feature_name); CvRect roi = cvGetImageROI(frontal); m_center = rect_center(roi); Allocate(pose_count, cvSize(roi.width, roi.height), frontal->nChannels); GenerateSamplesFast(frontal, pca_hr_avg, pca_hr_eigenvectors, pca_descriptors); } void OneWayDescriptor::InitializePCACoeffs(CvMat* avg, CvMat* eigenvectors) { for(int i = 0; i < m_pose_count; i++) { ProjectPCASample(m_samples[i], avg, eigenvectors, m_pca_coeffs[i]); } } void OneWayDescriptor::ProjectPCASample(IplImage* patch, CvMat* avg, CvMat* eigenvectors, CvMat* pca_coeffs) const { CvMat* patch_mat = ConvertImageToMatrix(patch); // CvMat eigenvectorsr; // cvGetSubRect(eigenvectors, &eigenvectorsr, cvRect(0, 0, eigenvectors->cols, pca_coeffs->cols)); CvMat* temp = cvCreateMat(1, eigenvectors->cols, CV_32FC1); cvProjectPCA(patch_mat, avg, eigenvectors, temp); CvMat temp1; cvGetSubRect(temp, &temp1, cvRect(0, 0, pca_coeffs->cols, 1)); cvCopy(&temp1, pca_coeffs); cvReleaseMat(&temp); cvReleaseMat(&patch_mat); } void OneWayDescriptor::EstimatePosePCA(CvArr* patch, int& pose_idx, float& distance, CvMat* avg, CvMat* eigenvectors) const { if(avg == 0) { // do not use pca if (!CV_IS_MAT(patch)) { EstimatePose((IplImage*)patch, pose_idx, distance); } else { } return; } CvRect roi={0,0,0,0}; if (!CV_IS_MAT(patch)) { roi = cvGetImageROI((IplImage*)patch); if(roi.width != GetPatchSize().width || roi.height != GetPatchSize().height) { cvResize(patch, m_input_patch); patch = m_input_patch; roi = cvGetImageROI((IplImage*)patch); } } CvMat* pca_coeffs = cvCreateMat(1, m_pca_dim_low, CV_32FC1); if (CV_IS_MAT(patch)) { cvCopy((CvMat*)patch, pca_coeffs); } else { IplImage* patch_32f = cvCreateImage(cvSize(roi.width, roi.height), IPL_DEPTH_32F, 1); double sum = cvSum(patch).val[0]; cvConvertScale(patch, patch_32f, 1.0f/sum); ProjectPCASample(patch_32f, avg, eigenvectors, pca_coeffs); cvReleaseImage(&patch_32f); } distance = 1e10; pose_idx = -1; for(int i = 0; i < m_pose_count; i++) { double dist = cvNorm(m_pca_coeffs[i], pca_coeffs); // float dist = 0; // float data1, data2; // //CvMat* pose_pca_coeffs = m_pca_coeffs[i]; // for (int x=0; x < pca_coeffs->width; x++) // for (int y =0 ; y < pca_coeffs->height; y++) // { // data1 = ((float*)(pca_coeffs->data.ptr + pca_coeffs->step*x))[y]; // data2 = ((float*)(m_pca_coeffs[i]->data.ptr + m_pca_coeffs[i]->step*x))[y]; // dist+=(data1-data2)*(data1-data2); // } ////#if 1 // for (int j = 0; j < m_pca_dim_low; j++) // { // dist += (pose_pca_coeffs->data.fl[j]- pca_coeffs->data.fl[j])*(pose_pca_coeffs->data.fl[j]- pca_coeffs->data.fl[j]); // } //#else // for (int j = 0; j <= m_pca_dim_low - 4; j += 4) // { // dist += (pose_pca_coeffs->data.fl[j]- pca_coeffs->data.fl[j])* // (pose_pca_coeffs->data.fl[j]- pca_coeffs->data.fl[j]); // dist += (pose_pca_coeffs->data.fl[j+1]- pca_coeffs->data.fl[j+1])* // (pose_pca_coeffs->data.fl[j+1]- pca_coeffs->data.fl[j+1]); // dist += (pose_pca_coeffs->data.fl[j+2]- pca_coeffs->data.fl[j+2])* // (pose_pca_coeffs->data.fl[j+2]- pca_coeffs->data.fl[j+2]); // dist += (pose_pca_coeffs->data.fl[j+3]- pca_coeffs->data.fl[j+3])* // (pose_pca_coeffs->data.fl[j+3]- pca_coeffs->data.fl[j+3]); // } //#endif if(dist < distance) { distance = (float)dist; pose_idx = i; } } cvReleaseMat(&pca_coeffs); } void OneWayDescriptor::EstimatePose(IplImage* patch, int& pose_idx, float& distance) const { distance = 1e10; pose_idx = -1; CvRect roi = cvGetImageROI(patch); IplImage* patch_32f = cvCreateImage(cvSize(roi.width, roi.height), IPL_DEPTH_32F, patch->nChannels); double sum = cvSum(patch).val[0]; cvConvertScale(patch, patch_32f, 1/sum); for(int i = 0; i < m_pose_count; i++) { if(m_samples[i]->width != patch_32f->width || m_samples[i]->height != patch_32f->height) { continue; } double dist = cvNorm(m_samples[i], patch_32f); //float dist = 0.0f; //float i1,i2; //for (int y = 0; yheight; y++) // for (int x = 0; x< patch_32f->width; x++) // { // i1 = ((float*)(m_samples[i]->imageData + m_samples[i]->widthStep*y))[x]; // i2 = ((float*)(patch_32f->imageData + patch_32f->widthStep*y))[x]; // dist+= (i1-i2)*(i1-i2); // } if(dist < distance) { distance = (float)dist; pose_idx = i; } #if 0 IplImage* img1 = cvCreateImage(cvSize(roi.width, roi.height), IPL_DEPTH_8U, 1); IplImage* img2 = cvCreateImage(cvSize(roi.width, roi.height), IPL_DEPTH_8U, 1); double maxval; cvMinMaxLoc(m_samples[i], 0, &maxval); cvConvertScale(m_samples[i], img1, 255.0/maxval); cvMinMaxLoc(patch_32f, 0, &maxval); cvConvertScale(patch_32f, img2, 255.0/maxval); cvNamedWindow("1", 1); cvShowImage("1", img1); cvNamedWindow("2", 1); cvShowImage("2", img2); printf("Distance = %f\n", dist); cvWaitKey(0); #endif } cvReleaseImage(&patch_32f); } void OneWayDescriptor::Save(const char* path) { for(int i = 0; i < m_pose_count; i++) { char buf[1024]; sprintf(buf, "%s/patch_%04d.jpg", path, i); IplImage* patch = cvCreateImage(cvSize(m_samples[i]->width, m_samples[i]->height), IPL_DEPTH_8U, m_samples[i]->nChannels); double maxval; cvMinMaxLoc(m_samples[i], 0, &maxval); cvConvertScale(m_samples[i], patch, 255/maxval); cvSaveImage(buf, patch); cvReleaseImage(&patch); } } void OneWayDescriptor::Write(CvFileStorage* fs, const char* name) { CvMat* mat = cvCreateMat(m_pose_count, m_samples[0]->width*m_samples[0]->height, CV_32FC1); // prepare data to write as a single matrix for(int i = 0; i < m_pose_count; i++) { for(int y = 0; y < m_samples[i]->height; y++) { for(int x = 0; x < m_samples[i]->width; x++) { float val = *((float*)(m_samples[i]->imageData + m_samples[i]->widthStep*y) + x); cvmSet(mat, i, y*m_samples[i]->width + x, val); } } } cvWrite(fs, name, mat); cvReleaseMat(&mat); } int OneWayDescriptor::ReadByName(const FileNode &parent, const char* name) { CvMat* mat = reinterpret_cast (parent[name].readObj ()); if(!mat) { return 0; } for(int i = 0; i < m_pose_count; i++) { for(int y = 0; y < m_samples[i]->height; y++) { for(int x = 0; x < m_samples[i]->width; x++) { float val = (float)cvmGet(mat, i, y*m_samples[i]->width + x); *((float*)(m_samples[i]->imageData + y*m_samples[i]->widthStep) + x) = val; } } } cvReleaseMat(&mat); return 1; } int OneWayDescriptor::ReadByName(CvFileStorage* fs, CvFileNode* parent, const char* name) { return ReadByName (FileNode (fs, parent), name); } IplImage* OneWayDescriptor::GetPatch(int index) { return m_samples[index]; } CvAffinePose OneWayDescriptor::GetPose(int index) const { return m_affine_poses[index]; } void FindOneWayDescriptor(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, int& desc_idx, int& pose_idx, float& distance, CvMat* avg, CvMat* eigenvectors) { desc_idx = -1; pose_idx = -1; distance = 1e10; //-------- //PCA_coeffs precalculating int m_pca_dim_low = descriptors[0].GetPCADimLow(); CvMat* pca_coeffs = cvCreateMat(1, m_pca_dim_low, CV_32FC1); int patch_width = descriptors[0].GetPatchSize().width; int patch_height = descriptors[0].GetPatchSize().height; if (avg) { CvRect _roi = cvGetImageROI((IplImage*)patch); IplImage* test_img = cvCreateImage(cvSize(patch_width,patch_height), IPL_DEPTH_8U, 1); if(_roi.width != patch_width|| _roi.height != patch_height) { cvResize(patch, test_img); _roi = cvGetImageROI(test_img); } else { cvCopy(patch,test_img); } IplImage* patch_32f = cvCreateImage(cvSize(_roi.width, _roi.height), IPL_DEPTH_32F, 1); double sum = cvSum(test_img).val[0]; cvConvertScale(test_img, patch_32f, 1.0f/sum); //ProjectPCASample(patch_32f, avg, eigenvectors, pca_coeffs); //Projecting PCA CvMat* patch_mat = ConvertImageToMatrix(patch_32f); CvMat* temp = cvCreateMat(1, eigenvectors->cols, CV_32FC1); cvProjectPCA(patch_mat, avg, eigenvectors, temp); CvMat temp1; cvGetSubRect(temp, &temp1, cvRect(0, 0, pca_coeffs->cols, 1)); cvCopy(&temp1, pca_coeffs); cvReleaseMat(&temp); cvReleaseMat(&patch_mat); //End of projecting cvReleaseImage(&patch_32f); cvReleaseImage(&test_img); } //-------- for(int i = 0; i < desc_count; i++) { int _pose_idx = -1; float _distance = 0; #if 0 descriptors[i].EstimatePose(patch, _pose_idx, _distance); #else if (!avg) { descriptors[i].EstimatePosePCA(patch, _pose_idx, _distance, avg, eigenvectors); } else { descriptors[i].EstimatePosePCA(pca_coeffs, _pose_idx, _distance, avg, eigenvectors); } #endif if(_distance < distance) { desc_idx = i; pose_idx = _pose_idx; distance = _distance; } } cvReleaseMat(&pca_coeffs); } #if defined(_KDTREE) void FindOneWayDescriptor(cv::flann::Index* m_pca_descriptors_tree, CvSize patch_size, int m_pca_dim_low, int m_pose_count, IplImage* patch, int& desc_idx, int& pose_idx, float& distance, CvMat* avg, CvMat* eigenvectors) { desc_idx = -1; pose_idx = -1; distance = 1e10; //-------- //PCA_coeffs precalculating CvMat* pca_coeffs = cvCreateMat(1, m_pca_dim_low, CV_32FC1); int patch_width = patch_size.width; int patch_height = patch_size.height; //if (avg) //{ CvRect _roi = cvGetImageROI((IplImage*)patch); IplImage* test_img = cvCreateImage(cvSize(patch_width,patch_height), IPL_DEPTH_8U, 1); if(_roi.width != patch_width|| _roi.height != patch_height) { cvResize(patch, test_img); _roi = cvGetImageROI(test_img); } else { cvCopy(patch,test_img); } IplImage* patch_32f = cvCreateImage(cvSize(_roi.width, _roi.height), IPL_DEPTH_32F, 1); float sum = cvSum(test_img).val[0]; cvConvertScale(test_img, patch_32f, 1.0f/sum); //ProjectPCASample(patch_32f, avg, eigenvectors, pca_coeffs); //Projecting PCA CvMat* patch_mat = ConvertImageToMatrix(patch_32f); CvMat* temp = cvCreateMat(1, eigenvectors->cols, CV_32FC1); cvProjectPCA(patch_mat, avg, eigenvectors, temp); CvMat temp1; cvGetSubRect(temp, &temp1, cvRect(0, 0, pca_coeffs->cols, 1)); cvCopy(&temp1, pca_coeffs); cvReleaseMat(&temp); cvReleaseMat(&patch_mat); //End of projecting cvReleaseImage(&patch_32f); cvReleaseImage(&test_img); // } //-------- //float* target = new float[m_pca_dim_low]; //::cvflann::KNNResultSet res(1,pca_coeffs->data.fl,m_pca_dim_low); //::cvflann::SearchParams params; //params.checks = -1; //int maxDepth = 1000000; //int neighbors_count = 1; //int* neighborsIdx = new int[neighbors_count]; //float* distances = new float[neighbors_count]; //if (m_pca_descriptors_tree->findNearest(pca_coeffs->data.fl,neighbors_count,maxDepth,neighborsIdx,0,distances) > 0) //{ // desc_idx = neighborsIdx[0] / m_pose_count; // pose_idx = neighborsIdx[0] % m_pose_count; // distance = distances[0]; //} //delete[] neighborsIdx; //delete[] distances; cv::Mat m_object(1, m_pca_dim_low, CV_32F); cv::Mat m_indices(1, 1, CV_32S); cv::Mat m_dists(1, 1, CV_32F); float* object_ptr = m_object.ptr(0); for (int i=0;idata.fl[i]; } m_pca_descriptors_tree->knnSearch(m_object, m_indices, m_dists, 1, cv::flann::SearchParams(-1) ); desc_idx = ((int*)(m_indices.ptr(0)))[0] / m_pose_count; pose_idx = ((int*)(m_indices.ptr(0)))[0] % m_pose_count; distance = ((float*)(m_dists.ptr(0)))[0]; // delete[] target; // for(int i = 0; i < desc_count; i++) // { // int _pose_idx = -1; // float _distance = 0; // //#if 0 // descriptors[i].EstimatePose(patch, _pose_idx, _distance); //#else // if (!avg) // { // descriptors[i].EstimatePosePCA(patch, _pose_idx, _distance, avg, eigenvectors); // } // else // { // descriptors[i].EstimatePosePCA(pca_coeffs, _pose_idx, _distance, avg, eigenvectors); // } //#endif // // if(_distance < distance) // { // desc_idx = i; // pose_idx = _pose_idx; // distance = _distance; // } // } cvReleaseMat(&pca_coeffs); } #endif //** void FindOneWayDescriptor(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, int n, std::vector& desc_idxs, std::vector& pose_idxs, std::vector& distances, CvMat* avg, CvMat* eigenvectors) { for (int i=0;icols, CV_32FC1); cvProjectPCA(patch_mat, avg, eigenvectors, temp); CvMat temp1; cvGetSubRect(temp, &temp1, cvRect(0, 0, pca_coeffs->cols, 1)); cvCopy(&temp1, pca_coeffs); cvReleaseMat(&temp); cvReleaseMat(&patch_mat); //End of projecting cvReleaseImage(&patch_32f); cvReleaseImage(&test_img); } //-------- for(int i = 0; i < desc_count; i++) { int _pose_idx = -1; float _distance = 0; #if 0 descriptors[i].EstimatePose(patch, _pose_idx, _distance); #else if (!avg) { descriptors[i].EstimatePosePCA(patch, _pose_idx, _distance, avg, eigenvectors); } else { descriptors[i].EstimatePosePCA(pca_coeffs, _pose_idx, _distance, avg, eigenvectors); } #endif for (int j=0;j j;k--) { desc_idxs[k] = desc_idxs[k-1]; pose_idxs[k] = pose_idxs[k-1]; distances[k] = distances[k-1]; } desc_idxs[j] = i; pose_idxs[j] = _pose_idx; distances[j] = _distance; break; } } } cvReleaseMat(&pca_coeffs); } void FindOneWayDescriptorEx(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, float scale_min, float scale_max, float scale_step, int& desc_idx, int& pose_idx, float& distance, float& scale, CvMat* avg, CvMat* eigenvectors) { CvSize patch_size = descriptors[0].GetPatchSize(); IplImage* input_patch; CvRect roi; input_patch= cvCreateImage(patch_size, IPL_DEPTH_8U, 1); roi = cvGetImageROI((IplImage*)patch); int _desc_idx, _pose_idx; float _distance; distance = 1e10; for(float cur_scale = scale_min; cur_scale < scale_max; cur_scale *= scale_step) { // printf("Scale = %f\n", cur_scale); CvRect roi_scaled = resize_rect(roi, cur_scale); cvSetImageROI(patch, roi_scaled); cvResize(patch, input_patch); #if 0 if(roi.x > 244 && roi.y < 200) { cvNamedWindow("1", 1); cvShowImage("1", input_patch); cvWaitKey(0); } #endif FindOneWayDescriptor(desc_count, descriptors, input_patch, _desc_idx, _pose_idx, _distance, avg, eigenvectors); if(_distance < distance) { distance = _distance; desc_idx = _desc_idx; pose_idx = _pose_idx; scale = cur_scale; } } cvSetImageROI((IplImage*)patch, roi); cvReleaseImage(&input_patch); } void FindOneWayDescriptorEx(int desc_count, const OneWayDescriptor* descriptors, IplImage* patch, float scale_min, float scale_max, float scale_step, int n, std::vector& desc_idxs, std::vector& pose_idxs, std::vector& distances, std::vector& scales, CvMat* avg, CvMat* eigenvectors) { CvSize patch_size = descriptors[0].GetPatchSize(); IplImage* input_patch; CvRect roi; input_patch= cvCreateImage(patch_size, IPL_DEPTH_8U, 1); roi = cvGetImageROI((IplImage*)patch); // float min_distance = 1e10; std::vector _desc_idxs; _desc_idxs.resize(n); std::vector _pose_idxs; _pose_idxs.resize(n); std::vector _distances; _distances.resize(n); for (int i=0;idepth == 32) { for(int y = 0; y < roi.height; y++) { for(int x = 0; x < roi.width; x++) { mat->data.fl[y*roi.width + x] = *((float*)(patch->imageData + (y + roi.y)*patch->widthStep) + x + roi.x); } } } else if(patch->depth == 8) { for(int y = 0; y < roi.height; y++) { for(int x = 0; x < roi.width; x++) { mat->data.fl[y*roi.width + x] = (float)(unsigned char)patch->imageData[(y + roi.y)*patch->widthStep + x + roi.x]; } } } else { printf("Image depth %d is not supported\n", patch->depth); return 0; } return mat; } OneWayDescriptorBase::OneWayDescriptorBase(CvSize patch_size, int pose_count, const char* train_path, const char* pca_config, const char* pca_hr_config, const char* pca_desc_config, int pyr_levels, int pca_dim_high, int pca_dim_low) : m_pca_dim_high(pca_dim_high), m_pca_dim_low(pca_dim_low), scale_min (0.7f), scale_max(1.5f), scale_step (1.2f) { #if defined(_KDTREE) m_pca_descriptors_matrix = 0; m_pca_descriptors_tree = 0; #endif // m_pca_descriptors_matrix = 0; m_patch_size = patch_size; m_pose_count = pose_count; m_pyr_levels = pyr_levels; m_poses = 0; m_transforms = 0; m_pca_avg = 0; m_pca_eigenvectors = 0; m_pca_hr_avg = 0; m_pca_hr_eigenvectors = 0; m_pca_descriptors = 0; m_descriptors = 0; if(train_path == 0 || strlen(train_path) == 0) { // skip pca loading return; } char pca_config_filename[1024]; sprintf(pca_config_filename, "%s/%s", train_path, pca_config); readPCAFeatures(pca_config_filename, &m_pca_avg, &m_pca_eigenvectors); if(pca_hr_config && strlen(pca_hr_config) > 0) { char pca_hr_config_filename[1024]; sprintf(pca_hr_config_filename, "%s/%s", train_path, pca_hr_config); readPCAFeatures(pca_hr_config_filename, &m_pca_hr_avg, &m_pca_hr_eigenvectors); } m_pca_descriptors = new OneWayDescriptor[m_pca_dim_high + 1]; #if !defined(_GH_REGIONS) if(pca_desc_config && strlen(pca_desc_config) > 0) // if(0) { //printf("Loading the descriptors..."); char pca_desc_config_filename[1024]; sprintf(pca_desc_config_filename, "%s/%s", train_path, pca_desc_config); LoadPCADescriptors(pca_desc_config_filename); //printf("done.\n"); } else { printf("Initializing the descriptors...\n"); InitializePoseTransforms(); CreatePCADescriptors(); SavePCADescriptors("pca_descriptors.yml"); } #endif //_GH_REGIONS // SavePCADescriptors("./pca_descriptors.yml"); } OneWayDescriptorBase::OneWayDescriptorBase(CvSize patch_size, int pose_count, const string &pca_filename, const string &train_path, const string &images_list, float _scale_min, float _scale_max, float _scale_step, int pyr_levels, int pca_dim_high, int pca_dim_low) : m_pca_dim_high(pca_dim_high), m_pca_dim_low(pca_dim_low), scale_min(_scale_min), scale_max(_scale_max), scale_step(_scale_step) { #if defined(_KDTREE) m_pca_descriptors_matrix = 0; m_pca_descriptors_tree = 0; #endif m_patch_size = patch_size; m_pose_count = pose_count; m_pyr_levels = pyr_levels; m_poses = 0; m_transforms = 0; m_pca_avg = 0; m_pca_eigenvectors = 0; m_pca_hr_avg = 0; m_pca_hr_eigenvectors = 0; m_pca_descriptors = 0; m_descriptors = 0; if (pca_filename.length() == 0) { return; } CvFileStorage* fs = cvOpenFileStorage(pca_filename.c_str(), NULL, CV_STORAGE_READ); if (fs != 0) { cvReleaseFileStorage(&fs); readPCAFeatures(pca_filename.c_str(), &m_pca_avg, &m_pca_eigenvectors, "_lr"); readPCAFeatures(pca_filename.c_str(), &m_pca_hr_avg, &m_pca_hr_eigenvectors, "_hr"); m_pca_descriptors = new OneWayDescriptor[m_pca_dim_high + 1]; #if !defined(_GH_REGIONS) LoadPCADescriptors(pca_filename.c_str()); #endif //_GH_REGIONS } else { GeneratePCA(train_path.c_str(), images_list.c_str()); m_pca_descriptors = new OneWayDescriptor[m_pca_dim_high + 1]; char pca_default_filename[1024]; sprintf(pca_default_filename, "%s/%s", train_path.c_str(), GetPCAFilename().c_str()); LoadPCADescriptors(pca_default_filename); } } void OneWayDescriptorBase::Read (const FileNode &fn) { clear (); m_pose_count = fn["poseCount"]; int patch_width = fn["patchWidth"]; int patch_height = fn["patchHeight"]; m_patch_size = cvSize (patch_width, patch_height); m_pyr_levels = fn["pyrLevels"]; m_pca_dim_high = fn["pcaDimHigh"]; m_pca_dim_low = fn["pcaDimLow"]; scale_min = fn["minScale"]; scale_max = fn["maxScale"]; scale_step = fn["stepScale"]; LoadPCAall (fn); } void OneWayDescriptorBase::LoadPCAall (const FileNode &fn) { readPCAFeatures(fn, &m_pca_avg, &m_pca_eigenvectors, "_lr"); readPCAFeatures(fn, &m_pca_hr_avg, &m_pca_hr_eigenvectors, "_hr"); m_pca_descriptors = new OneWayDescriptor[m_pca_dim_high + 1]; #if !defined(_GH_REGIONS) LoadPCADescriptors(fn); #endif //_GH_REGIONS } OneWayDescriptorBase::~OneWayDescriptorBase() { cvReleaseMat(&m_pca_avg); cvReleaseMat(&m_pca_eigenvectors); if(m_pca_hr_eigenvectors) { delete[] m_pca_descriptors; cvReleaseMat(&m_pca_hr_avg); cvReleaseMat(&m_pca_hr_eigenvectors); } if(m_descriptors) delete []m_descriptors; if(m_poses) delete []m_poses; if (m_transforms) { for(int i = 0; i < m_pose_count; i++) { cvReleaseMat(&m_transforms[i]); } delete []m_transforms; } #if defined(_KDTREE) if (m_pca_descriptors_matrix) { cvReleaseMat(&m_pca_descriptors_matrix); } if (m_pca_descriptors_tree) { delete m_pca_descriptors_tree; } #endif } void OneWayDescriptorBase::clear(){ if (m_descriptors) { delete []m_descriptors; m_descriptors = 0; } #if defined(_KDTREE) if (m_pca_descriptors_matrix) { cvReleaseMat(&m_pca_descriptors_matrix); m_pca_descriptors_matrix = 0; } if (m_pca_descriptors_tree) { delete m_pca_descriptors_tree; m_pca_descriptors_tree = 0; } #endif } void OneWayDescriptorBase::InitializePoses() { m_poses = new CvAffinePose[m_pose_count]; for(int i = 0; i < m_pose_count; i++) { m_poses[i] = GenRandomAffinePose(); } } void OneWayDescriptorBase::InitializeTransformsFromPoses() { m_transforms = new CvMat*[m_pose_count]; for(int i = 0; i < m_pose_count; i++) { m_transforms[i] = cvCreateMat(2, 3, CV_32FC1); GenerateAffineTransformFromPose(cvSize(m_patch_size.width*2, m_patch_size.height*2), m_poses[i], m_transforms[i]); } } void OneWayDescriptorBase::InitializePoseTransforms() { InitializePoses(); InitializeTransformsFromPoses(); } void OneWayDescriptorBase::InitializeDescriptor(int desc_idx, IplImage* train_image, const KeyPoint& keypoint, const char* feature_label) { // TBD add support for octave != 0 CvPoint center = keypoint.pt; CvRect roi = cvRect(center.x - m_patch_size.width/2, center.y - m_patch_size.height/2, m_patch_size.width, m_patch_size.height); cvResetImageROI(train_image); roi = fit_rect_fixedsize(roi, train_image); cvSetImageROI(train_image, roi); if(roi.width != m_patch_size.width || roi.height != m_patch_size.height) { return; } InitializeDescriptor(desc_idx, train_image, feature_label); cvResetImageROI(train_image); } void OneWayDescriptorBase::InitializeDescriptor(int desc_idx, IplImage* train_image, const char* feature_label) { m_descriptors[desc_idx].SetPCADimHigh(m_pca_dim_high); m_descriptors[desc_idx].SetPCADimLow(m_pca_dim_low); m_descriptors[desc_idx].SetTransforms(m_poses, m_transforms); if(!m_pca_hr_eigenvectors) { m_descriptors[desc_idx].Initialize(m_pose_count, train_image, feature_label); } else { m_descriptors[desc_idx].InitializeFast(m_pose_count, train_image, feature_label, m_pca_hr_avg, m_pca_hr_eigenvectors, m_pca_descriptors); } if(m_pca_avg) { m_descriptors[desc_idx].InitializePCACoeffs(m_pca_avg, m_pca_eigenvectors); } } void OneWayDescriptorBase::FindDescriptor(IplImage* src, cv::Point2f pt, int& desc_idx, int& pose_idx, float& distance) const { CvRect roi = cvRect(cvRound(pt.x - m_patch_size.width/4), cvRound(pt.y - m_patch_size.height/4), m_patch_size.width/2, m_patch_size.height/2); cvSetImageROI(src, roi); FindDescriptor(src, desc_idx, pose_idx, distance); cvResetImageROI(src); } void OneWayDescriptorBase::FindDescriptor(IplImage* patch, int& desc_idx, int& pose_idx, float& distance, float* _scale, float* scale_ranges) const { #if 0 ::FindOneWayDescriptor(m_train_feature_count, m_descriptors, patch, desc_idx, pose_idx, distance, m_pca_avg, m_pca_eigenvectors); #else float min = scale_min; float max = scale_max; float step = scale_step; if (scale_ranges) { min = scale_ranges[0]; max = scale_ranges[1]; } float scale = 1.0f; #if !defined(_KDTREE) cv::FindOneWayDescriptorEx(m_train_feature_count, m_descriptors, patch, min, max, step, desc_idx, pose_idx, distance, scale, m_pca_avg, m_pca_eigenvectors); #else cv::FindOneWayDescriptorEx(m_pca_descriptors_tree, m_descriptors[0].GetPatchSize(), m_descriptors[0].GetPCADimLow(), m_pose_count, patch, min, max, step, desc_idx, pose_idx, distance, scale, m_pca_avg, m_pca_eigenvectors); #endif if (_scale) *_scale = scale; #endif } void OneWayDescriptorBase::FindDescriptor(IplImage* patch, int n, std::vector& desc_idxs, std::vector& pose_idxs, std::vector& distances, std::vector& _scales, float* scale_ranges) const { float min = scale_min; float max = scale_max; float step = scale_step; if (scale_ranges) { min = scale_ranges[0]; max = scale_ranges[1]; } distances.resize(n); _scales.resize(n); desc_idxs.resize(n); pose_idxs.resize(n); /*float scales = 1.0f;*/ cv::FindOneWayDescriptorEx(m_train_feature_count, m_descriptors, patch, min, max, step ,n, desc_idxs, pose_idxs, distances, _scales, m_pca_avg, m_pca_eigenvectors); } void OneWayDescriptorBase::SetPCAHigh(CvMat* avg, CvMat* eigenvectors) { m_pca_hr_avg = cvCloneMat(avg); m_pca_hr_eigenvectors = cvCloneMat(eigenvectors); } void OneWayDescriptorBase::SetPCALow(CvMat* avg, CvMat* eigenvectors) { m_pca_avg = cvCloneMat(avg); m_pca_eigenvectors = cvCloneMat(eigenvectors); } void OneWayDescriptorBase::AllocatePCADescriptors() { m_pca_descriptors = new OneWayDescriptor[m_pca_dim_high + 1]; for(int i = 0; i < m_pca_dim_high + 1; i++) { m_pca_descriptors[i].SetPCADimHigh(m_pca_dim_high); m_pca_descriptors[i].SetPCADimLow(m_pca_dim_low); } } void OneWayDescriptorBase::CreatePCADescriptors() { if(m_pca_descriptors == 0) { AllocatePCADescriptors(); } IplImage* frontal = cvCreateImage(m_patch_size, IPL_DEPTH_32F, 1); eigenvector2image(m_pca_hr_avg, frontal); m_pca_descriptors[0].SetTransforms(m_poses, m_transforms); m_pca_descriptors[0].Initialize(m_pose_count, frontal, "", 0); for(int j = 0; j < m_pca_dim_high; j++) { CvMat eigenvector; cvGetSubRect(m_pca_hr_eigenvectors, &eigenvector, cvRect(0, j, m_pca_hr_eigenvectors->cols, 1)); eigenvector2image(&eigenvector, frontal); m_pca_descriptors[j + 1].SetTransforms(m_poses, m_transforms); m_pca_descriptors[j + 1].Initialize(m_pose_count, frontal, "", 0); printf("Created descriptor for PCA component %d\n", j); } cvReleaseImage(&frontal); } int OneWayDescriptorBase::LoadPCADescriptors(const char* filename) { FileStorage fs = FileStorage (filename, FileStorage::READ); if(!fs.isOpened ()) { printf("File %s not found...\n", filename); return 0; } LoadPCADescriptors (fs.root ()); printf("Successfully read %d pca components\n", m_pca_dim_high); fs.release (); return 1; } int OneWayDescriptorBase::LoadPCADescriptors(const FileNode &fn) { // read affine poses // FileNode* node = cvGetFileNodeByName(fs, 0, "affine poses"); CvMat* poses = reinterpret_cast (fn["affine_poses"].readObj ()); if (poses == 0) { poses = reinterpret_cast (fn["affine poses"].readObj ()); if (poses == 0) return 0; } if(m_poses) { delete m_poses; } m_poses = new CvAffinePose[m_pose_count]; for(int i = 0; i < m_pose_count; i++) { m_poses[i].phi = (float)cvmGet(poses, i, 0); m_poses[i].theta = (float)cvmGet(poses, i, 1); m_poses[i].lambda1 = (float)cvmGet(poses, i, 2); m_poses[i].lambda2 = (float)cvmGet(poses, i, 3); } cvReleaseMat(&poses); // now initialize pose transforms InitializeTransformsFromPoses(); m_pca_dim_high = (int) fn["pca_components_number"]; if (m_pca_dim_high == 0) { m_pca_dim_high = (int) fn["pca components number"]; } if(m_pca_descriptors) { delete []m_pca_descriptors; } AllocatePCADescriptors(); for(int i = 0; i < m_pca_dim_high + 1; i++) { m_pca_descriptors[i].Allocate(m_pose_count, m_patch_size, 1); m_pca_descriptors[i].SetTransforms(m_poses, m_transforms); char buf[1024]; sprintf(buf, "descriptor_for_pca_component_%d", i); if (! m_pca_descriptors[i].ReadByName(fn, buf)) { char buf[1024]; sprintf(buf, "descriptor for pca component %d", i); m_pca_descriptors[i].ReadByName(fn, buf); } } return 1; } void savePCAFeatures(FileStorage &fs, const char* postfix, CvMat* avg, CvMat* eigenvectors) { char buf[1024]; sprintf(buf, "avg_%s", postfix); fs.writeObj(buf, avg); sprintf(buf, "eigenvectors_%s", postfix); fs.writeObj(buf, eigenvectors); } void calcPCAFeatures(vector& patches, FileStorage &fs, const char* postfix, CvMat** avg, CvMat** eigenvectors) { int width = patches[0]->width; int height = patches[0]->height; int length = width * height; int patch_count = (int)patches.size(); CvMat* data = cvCreateMat(patch_count, length, CV_32FC1); *avg = cvCreateMat(1, length, CV_32FC1); CvMat* eigenvalues = cvCreateMat(1, length, CV_32FC1); *eigenvectors = cvCreateMat(length, length, CV_32FC1); for (int i = 0; i < patch_count; i++) { float nf = (float)(1./cvSum(patches[i]).val[0]); for (int y = 0; y < height; y++) { for (int x = 0; x < width; x++) { *((float*)(data->data.ptr + data->step * i) + y * width + x) = (unsigned char)patches[i]->imageData[y * patches[i]->widthStep + x] * nf; } } } //printf("Calculating PCA..."); cvCalcPCA(data, *avg, eigenvalues, *eigenvectors, CV_PCA_DATA_AS_ROW); //printf("done\n"); // save pca data savePCAFeatures(fs, postfix, *avg, *eigenvectors); cvReleaseMat(&data); cvReleaseMat(&eigenvalues); } void extractPatches (IplImage *img, vector& patches, CvSize patch_size) { vector features; Ptr surf_extractor = FeatureDetector::create("SURF"); if( surf_extractor.empty() ) CV_Error(CV_StsNotImplemented, "OpenCV was built without SURF support"); surf_extractor->set("hessianThreshold", 1.0); //printf("Extracting SURF features..."); surf_extractor->detect(Mat(img), features); //printf("done\n"); for (int j = 0; j < (int)features.size(); j++) { int patch_width = patch_size.width; int patch_height = patch_size.height; CvPoint center = features[j].pt; CvRect roi = cvRect(center.x - patch_width / 2, center.y - patch_height / 2, patch_width, patch_height); cvSetImageROI(img, roi); roi = cvGetImageROI(img); if (roi.width != patch_width || roi.height != patch_height) { continue; } IplImage* patch = cvCreateImage(cvSize(patch_width, patch_height), IPL_DEPTH_8U, 1); cvCopy(img, patch); patches.push_back(patch); cvResetImageROI(img); } //printf("Completed file, extracted %d features\n", (int)features.size()); } /* void loadPCAFeatures(const FileNode &fn, vector& patches, CvSize patch_size) { FileNodeIterator begin = fn.begin(); for (FileNodeIterator i = fn.begin(); i != fn.end(); i++) { IplImage *img = reinterpret_cast ((*i).readObj()); extractPatches (img, patches, patch_size); cvReleaseImage(&img); } } */ void loadPCAFeatures(const char* path, const char* images_list, vector& patches, CvSize patch_size) { char images_filename[1024]; sprintf(images_filename, "%s/%s", path, images_list); FILE *pFile = fopen(images_filename, "r"); if (pFile == 0) { printf("Cannot open images list file %s\n", images_filename); return; } while (!feof(pFile)) { char imagename[1024]; if (fscanf(pFile, "%s", imagename) <= 0) { break; } char filename[1024]; sprintf(filename, "%s/%s", path, imagename); //printf("Reading image %s...", filename); IplImage* img = cvLoadImage(filename, CV_LOAD_IMAGE_GRAYSCALE); //printf("done\n"); extractPatches (img, patches, patch_size); cvReleaseImage(&img); } fclose(pFile); } void generatePCAFeatures(const char* path, const char* img_filename, FileStorage& fs, const char* postfix, CvSize patch_size, CvMat** avg, CvMat** eigenvectors) { vector patches; loadPCAFeatures(path, img_filename, patches, patch_size); calcPCAFeatures(patches, fs, postfix, avg, eigenvectors); } /* void generatePCAFeatures(const FileNode &fn, const char* postfix, CvSize patch_size, CvMat** avg, CvMat** eigenvectors) { vector patches; loadPCAFeatures(fn, patches, patch_size); calcPCAFeatures(patches, fs, postfix, avg, eigenvectors); } void OneWayDescriptorBase::GeneratePCA(const FileNode &fn, int pose_count) { generatePCAFeatures(fn, "hr", m_patch_size, &m_pca_hr_avg, &m_pca_hr_eigenvectors); generatePCAFeatures(fn, "lr", cvSize(m_patch_size.width / 2, m_patch_size.height / 2), &m_pca_avg, &m_pca_eigenvectors); OneWayDescriptorBase descriptors(m_patch_size, pose_count); descriptors.SetPCAHigh(m_pca_hr_avg, m_pca_hr_eigenvectors); descriptors.SetPCALow(m_pca_avg, m_pca_eigenvectors); printf("Calculating %d PCA descriptors (you can grab a coffee, this will take a while)...\n", descriptors.GetPCADimHigh()); descriptors.InitializePoseTransforms(); descriptors.CreatePCADescriptors(); descriptors.SavePCADescriptors(*fs); } */ void OneWayDescriptorBase::GeneratePCA(const char* img_path, const char* images_list, int pose_count) { char pca_filename[1024]; sprintf(pca_filename, "%s/%s", img_path, GetPCAFilename().c_str()); FileStorage fs = FileStorage(pca_filename, FileStorage::WRITE); generatePCAFeatures(img_path, images_list, fs, "hr", m_patch_size, &m_pca_hr_avg, &m_pca_hr_eigenvectors); generatePCAFeatures(img_path, images_list, fs, "lr", cvSize(m_patch_size.width / 2, m_patch_size.height / 2), &m_pca_avg, &m_pca_eigenvectors); OneWayDescriptorBase descriptors(m_patch_size, pose_count); descriptors.SetPCAHigh(m_pca_hr_avg, m_pca_hr_eigenvectors); descriptors.SetPCALow(m_pca_avg, m_pca_eigenvectors); printf("Calculating %d PCA descriptors (you can grab a coffee, this will take a while)...\n", descriptors.GetPCADimHigh()); descriptors.InitializePoseTransforms(); descriptors.CreatePCADescriptors(); descriptors.SavePCADescriptors(*fs); fs.release(); } void OneWayDescriptorBase::Write (FileStorage &fs) const { fs << "poseCount" << m_pose_count; fs << "patchWidth" << m_patch_size.width; fs << "patchHeight" << m_patch_size.height; fs << "minScale" << scale_min; fs << "maxScale" << scale_max; fs << "stepScale" << scale_step; fs << "pyrLevels" << m_pyr_levels; fs << "pcaDimHigh" << m_pca_dim_high; fs << "pcaDimLow" << m_pca_dim_low; SavePCAall (fs); } void OneWayDescriptorBase::SavePCAall (FileStorage &fs) const { savePCAFeatures(fs, "hr", m_pca_hr_avg, m_pca_hr_eigenvectors); savePCAFeatures(fs, "lr", m_pca_avg, m_pca_eigenvectors); SavePCADescriptors(*fs); } void OneWayDescriptorBase::SavePCADescriptors(const char* filename) { CvMemStorage* storage = cvCreateMemStorage(); CvFileStorage* fs = cvOpenFileStorage(filename, storage, CV_STORAGE_WRITE); SavePCADescriptors (fs); cvReleaseMemStorage(&storage); cvReleaseFileStorage(&fs); } void OneWayDescriptorBase::SavePCADescriptors(CvFileStorage *fs) const { cvWriteInt(fs, "pca_components_number", m_pca_dim_high); cvWriteComment( fs, "The first component is the average Vector, so the total number of components is + 1", 0); cvWriteInt(fs, "patch_width", m_patch_size.width); cvWriteInt(fs, "patch_height", m_patch_size.height); // pack the affine transforms into a single CvMat and write them CvMat* poses = cvCreateMat(m_pose_count, 4, CV_32FC1); for (int i = 0; i < m_pose_count; i++) { cvmSet(poses, i, 0, m_poses[i].phi); cvmSet(poses, i, 1, m_poses[i].theta); cvmSet(poses, i, 2, m_poses[i].lambda1); cvmSet(poses, i, 3, m_poses[i].lambda2); } cvWrite(fs, "affine_poses", poses); cvReleaseMat(&poses); for (int i = 0; i < m_pca_dim_high + 1; i++) { char buf[1024]; sprintf(buf, "descriptor_for_pca_component_%d", i); m_pca_descriptors[i].Write(fs, buf); } } void OneWayDescriptorBase::Allocate(int train_feature_count) { m_train_feature_count = train_feature_count; m_descriptors = new OneWayDescriptor[m_train_feature_count]; for(int i = 0; i < m_train_feature_count; i++) { m_descriptors[i].SetPCADimHigh(m_pca_dim_high); m_descriptors[i].SetPCADimLow(m_pca_dim_low); } } void OneWayDescriptorBase::InitializeDescriptors(IplImage* train_image, const vector& features, const char* feature_label, int desc_start_idx) { for(int i = 0; i < (int)features.size(); i++) { InitializeDescriptor(desc_start_idx + i, train_image, features[i], feature_label); } cvResetImageROI(train_image); #if defined(_KDTREE) ConvertDescriptorsArrayToTree(); #endif } void OneWayDescriptorBase::CreateDescriptorsFromImage(IplImage* src, const std::vector& features) { m_train_feature_count = (int)features.size(); m_descriptors = new OneWayDescriptor[m_train_feature_count]; InitializeDescriptors(src, features); } #if defined(_KDTREE) void OneWayDescriptorBase::ConvertDescriptorsArrayToTree() { int n = this->GetDescriptorCount(); if (n<1) return; int pca_dim_low = this->GetDescriptor(0)->GetPCADimLow(); //if (!m_pca_descriptors_matrix) // m_pca_descriptors_matrix = new ::cvflann::Matrix(n*m_pose_count,pca_dim_low); //else //{ // if ((m_pca_descriptors_matrix->cols != pca_dim_low)&&(m_pca_descriptors_matrix->rows != n*m_pose_count)) // { // delete m_pca_descriptors_matrix; // m_pca_descriptors_matrix = new ::cvflann::Matrix(n*m_pose_count,pca_dim_low); // } //} m_pca_descriptors_matrix = cvCreateMat(n*m_pose_count,pca_dim_low,CV_32FC1); for (int i=0;idata.fl[(i*m_pose_count+j)*m_pca_dim_low + k] = pca_coeffs[j]->data.fl[k]; } } } cv::Mat pca_descriptors_mat(m_pca_descriptors_matrix,false); //::cvflann::KDTreeIndexParams params; //params.trees = 1; //m_pca_descriptors_tree = new KDTree(pca_descriptors_mat); m_pca_descriptors_tree = new cv::flann::Index(pca_descriptors_mat,cv::flann::KDTreeIndexParams(1)); //cvReleaseMat(&m_pca_descriptors_matrix); //m_pca_descriptors_tree->buildIndex(); } #endif void OneWayDescriptorObject::Allocate(int train_feature_count, int object_feature_count) { OneWayDescriptorBase::Allocate(train_feature_count); m_object_feature_count = object_feature_count; m_part_id = new int[m_object_feature_count]; } void OneWayDescriptorObject::InitializeObjectDescriptors(IplImage* train_image, const vector& features, const char* feature_label, int desc_start_idx, float scale, int is_background) { InitializeDescriptors(train_image, features, feature_label, desc_start_idx); for(int i = 0; i < (int)features.size(); i++) { CvPoint center = features[i].pt; if(!is_background) { // remember descriptor part id CvPoint center_scaled = cvPoint(round(center.x*scale), round(center.y*scale)); m_part_id[i + desc_start_idx] = MatchPointToPart(center_scaled); } } cvResetImageROI(train_image); } int OneWayDescriptorObject::IsDescriptorObject(int desc_idx) const { return desc_idx < m_object_feature_count ? 1 : 0; } int OneWayDescriptorObject::MatchPointToPart(CvPoint pt) const { int idx = -1; const int max_dist = 10; for(int i = 0; i < (int)m_train_features.size(); i++) { if(norm(Point2f(pt) - m_train_features[i].pt) < max_dist) { idx = i; break; } } return idx; } int OneWayDescriptorObject::GetDescriptorPart(int desc_idx) const { // return MatchPointToPart(GetDescriptor(desc_idx)->GetCenter()); return desc_idx < m_object_feature_count ? m_part_id[desc_idx] : -1; } OneWayDescriptorObject::OneWayDescriptorObject(CvSize patch_size, int pose_count, const char* train_path, const char* pca_config, const char* pca_hr_config, const char* pca_desc_config, int pyr_levels) : OneWayDescriptorBase(patch_size, pose_count, train_path, pca_config, pca_hr_config, pca_desc_config, pyr_levels) { m_part_id = 0; } OneWayDescriptorObject::OneWayDescriptorObject(CvSize patch_size, int pose_count, const string &pca_filename, const string &train_path, const string &images_list, float _scale_min, float _scale_max, float _scale_step, int pyr_levels) : OneWayDescriptorBase(patch_size, pose_count, pca_filename, train_path, images_list, _scale_min, _scale_max, _scale_step, pyr_levels) { m_part_id = 0; } OneWayDescriptorObject::~OneWayDescriptorObject() { if (m_part_id) delete []m_part_id; } vector OneWayDescriptorObject::_GetLabeledFeatures() const { vector features; for(size_t i = 0; i < m_train_features.size(); i++) { features.push_back(m_train_features[i]); } return features; } void eigenvector2image(CvMat* eigenvector, IplImage* img) { CvRect roi = cvGetImageROI(img); if(img->depth == 32) { for(int y = 0; y < roi.height; y++) { for(int x = 0; x < roi.width; x++) { float val = (float)cvmGet(eigenvector, 0, roi.width*y + x); *((float*)(img->imageData + (roi.y + y)*img->widthStep) + roi.x + x) = val; } } } else { for(int y = 0; y < roi.height; y++) { for(int x = 0; x < roi.width; x++) { float val = (float)cvmGet(eigenvector, 0, roi.width*y + x); img->imageData[(roi.y + y)*img->widthStep + roi.x + x] = (unsigned char)val; } } } } void readPCAFeatures(const char* filename, CvMat** avg, CvMat** eigenvectors, const char* postfix) { FileStorage fs = FileStorage(filename, FileStorage::READ); if (!fs.isOpened ()) { printf("Cannot open file %s! Exiting!", filename); } readPCAFeatures (fs.root (), avg, eigenvectors, postfix); fs.release (); } void readPCAFeatures(const FileNode &fn, CvMat** avg, CvMat** eigenvectors, const char* postfix) { std::string str = std::string ("avg") + postfix; CvMat* _avg = reinterpret_cast (fn[str].readObj()); if (_avg != 0) { *avg = cvCloneMat(_avg); cvReleaseMat(&_avg); } str = std::string ("eigenvectors") + postfix; CvMat* _eigenvectors = reinterpret_cast (fn[str].readObj()); if (_eigenvectors != 0) { *eigenvectors = cvCloneMat(_eigenvectors); cvReleaseMat(&_eigenvectors); } } /****************************************************************************************\ * OneWayDescriptorMatcher * \****************************************************************************************/ OneWayDescriptorMatcher::Params::Params( int _poseCount, Size _patchSize, string _pcaFilename, string _trainPath, string _trainImagesList, float _minScale, float _maxScale, float _stepScale ) : poseCount(_poseCount), patchSize(_patchSize), pcaFilename(_pcaFilename), trainPath(_trainPath), trainImagesList(_trainImagesList), minScale(_minScale), maxScale(_maxScale), stepScale(_stepScale) {} OneWayDescriptorMatcher::OneWayDescriptorMatcher( const Params& _params) { initialize(_params); } OneWayDescriptorMatcher::~OneWayDescriptorMatcher() {} void OneWayDescriptorMatcher::initialize( const Params& _params, const Ptr& _base ) { clear(); if( _base.empty() ) base = _base; params = _params; } void OneWayDescriptorMatcher::clear() { GenericDescriptorMatcher::clear(); prevTrainCount = 0; if( !base.empty() ) base->clear(); } void OneWayDescriptorMatcher::train() { if( base.empty() || prevTrainCount < (int)trainPointCollection.keypointCount() ) { base = new OneWayDescriptorObject( params.patchSize, params.poseCount, params.pcaFilename, params.trainPath, params.trainImagesList, params.minScale, params.maxScale, params.stepScale ); base->Allocate( (int)trainPointCollection.keypointCount() ); prevTrainCount = (int)trainPointCollection.keypointCount(); const vector >& points = trainPointCollection.getKeypoints(); int count = 0; for( size_t i = 0; i < points.size(); i++ ) { IplImage _image = trainPointCollection.getImage((int)i); for( size_t j = 0; j < points[i].size(); j++ ) base->InitializeDescriptor( count++, &_image, points[i][j], "" ); } #if defined(_KDTREE) base->ConvertDescriptorsArrayToTree(); #endif } } bool OneWayDescriptorMatcher::isMaskSupported() { return false; } void OneWayDescriptorMatcher::knnMatchImpl( const Mat& queryImage, vector& queryKeypoints, vector >& matches, int knn, const vector& /*masks*/, bool /*compactResult*/ ) { train(); CV_Assert( knn == 1 ); // knn > 1 unsupported because of bug in OneWayDescriptorBase for this case matches.resize( queryKeypoints.size() ); IplImage _qimage = queryImage; for( size_t i = 0; i < queryKeypoints.size(); i++ ) { int descIdx = -1, poseIdx = -1; float distance; base->FindDescriptor( &_qimage, queryKeypoints[i].pt, descIdx, poseIdx, distance ); matches[i].push_back( DMatch((int)i, descIdx, distance) ); } } void OneWayDescriptorMatcher::radiusMatchImpl( const Mat& queryImage, vector& queryKeypoints, vector >& matches, float maxDistance, const vector& /*masks*/, bool /*compactResult*/ ) { train(); matches.resize( queryKeypoints.size() ); IplImage _qimage = queryImage; for( size_t i = 0; i < queryKeypoints.size(); i++ ) { int descIdx = -1, poseIdx = -1; float distance; base->FindDescriptor( &_qimage, queryKeypoints[i].pt, descIdx, poseIdx, distance ); if( distance < maxDistance ) matches[i].push_back( DMatch((int)i, descIdx, distance) ); } } void OneWayDescriptorMatcher::read( const FileNode &fn ) { base = new OneWayDescriptorObject( params.patchSize, params.poseCount, string (), string (), string (), params.minScale, params.maxScale, params.stepScale ); base->Read (fn); } void OneWayDescriptorMatcher::write( FileStorage& fs ) const { base->Write (fs); } bool OneWayDescriptorMatcher::empty() const { return base.empty() || base->empty(); } Ptr OneWayDescriptorMatcher::clone( bool emptyTrainData ) const { OneWayDescriptorMatcher* matcher = new OneWayDescriptorMatcher( params ); if( !emptyTrainData ) { CV_Error( CV_StsNotImplemented, "deep clone functionality is not implemented, because " "OneWayDescriptorBase has not copy constructor or clone method "); //matcher->base; matcher->params = params; matcher->prevTrainCount = prevTrainCount; matcher->trainPointCollection = trainPointCollection; } return matcher; } }