1 /*M/////////////////////////////////////////////////////////////////////////////////////// 2 // 3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. 4 // 5 // By downloading, copying, installing or using the software you agree to this license. 6 // If you do not agree to this license, do not download, install, 7 // copy or use the software. 8 // 9 // 10 // License Agreement 11 // For Open Source Computer Vision Library 12 // 13 // Copyright (C) 2000-2008, Intel Corporation, all rights reserved. 14 // Copyright (C) 2009, Willow Garage Inc., all rights reserved. 15 // Third party copyrights are property of their respective owners. 16 // 17 // Redistribution and use in source and binary forms, with or without modification, 18 // are permitted provided that the following conditions are met: 19 // 20 // * Redistribution's of source code must retain the above copyright notice, 21 // this list of conditions and the following disclaimer. 22 // 23 // * Redistribution's in binary form must reproduce the above copyright notice, 24 // this list of conditions and the following disclaimer in the documentation 25 // and/or other materials provided with the distribution. 26 // 27 // * The name of the copyright holders may not be used to endorse or promote products 28 // derived from this software without specific prior written permission. 29 // 30 // This software is provided by the copyright holders and contributors "as is" and 31 // any express or implied warranties, including, but not limited to, the implied 32 // warranties of merchantability and fitness for a particular purpose are disclaimed. 33 // In no event shall the Intel Corporation or contributors be liable for any direct, 34 // indirect, incidental, special, exemplary, or consequential damages 35 // (including, but not limited to, procurement of substitute goods or services; 36 // loss of use, data, or profits; or business interruption) however caused 37 // and on any theory of liability, whether in contract, strict liability, 38 // or tort (including negligence or otherwise) arising in any way out of 39 // the use of this software, even if advised of the possibility of such damage. 40 // 41 //M*/ 42 43 #include "precomp.hpp" 44 45 using namespace cv; 46 using namespace cv::detail; 47 using namespace cv::cuda; 48 49 #ifdef HAVE_OPENCV_XFEATURES2D 50 #include "opencv2/xfeatures2d.hpp" 51 using xfeatures2d::SURF; 52 #endif 53 54 namespace { 55 56 struct DistIdxPair 57 { 58 bool operator<(const DistIdxPair &other) const { return dist < other.dist; } 59 double dist; 60 int idx; 61 }; 62 63 64 struct MatchPairsBody : ParallelLoopBody 65 { 66 MatchPairsBody(FeaturesMatcher &_matcher, const std::vector<ImageFeatures> &_features, 67 std::vector<MatchesInfo> &_pairwise_matches, std::vector<std::pair<int,int> > &_near_pairs) 68 : matcher(_matcher), features(_features), 69 pairwise_matches(_pairwise_matches), near_pairs(_near_pairs) {} 70 71 void operator ()(const Range &r) const 72 { 73 const int num_images = static_cast<int>(features.size()); 74 for (int i = r.start; i < r.end; ++i) 75 { 76 int from = near_pairs[i].first; 77 int to = near_pairs[i].second; 78 int pair_idx = from*num_images + to; 79 80 matcher(features[from], features[to], pairwise_matches[pair_idx]); 81 pairwise_matches[pair_idx].src_img_idx = from; 82 pairwise_matches[pair_idx].dst_img_idx = to; 83 84 size_t dual_pair_idx = to*num_images + from; 85 86 pairwise_matches[dual_pair_idx] = pairwise_matches[pair_idx]; 87 pairwise_matches[dual_pair_idx].src_img_idx = to; 88 pairwise_matches[dual_pair_idx].dst_img_idx = from; 89 90 if (!pairwise_matches[pair_idx].H.empty()) 91 pairwise_matches[dual_pair_idx].H = pairwise_matches[pair_idx].H.inv(); 92 93 for (size_t j = 0; j < pairwise_matches[dual_pair_idx].matches.size(); ++j) 94 std::swap(pairwise_matches[dual_pair_idx].matches[j].queryIdx, 95 pairwise_matches[dual_pair_idx].matches[j].trainIdx); 96 LOG("."); 97 } 98 } 99 100 FeaturesMatcher &matcher; 101 const std::vector<ImageFeatures> &features; 102 std::vector<MatchesInfo> &pairwise_matches; 103 std::vector<std::pair<int,int> > &near_pairs; 104 105 private: 106 void operator =(const MatchPairsBody&); 107 }; 108 109 110 ////////////////////////////////////////////////////////////////////////////// 111 112 typedef std::set<std::pair<int,int> > MatchesSet; 113 114 // These two classes are aimed to find features matches only, not to 115 // estimate homography 116 117 class CpuMatcher : public FeaturesMatcher 118 { 119 public: 120 CpuMatcher(float match_conf) : FeaturesMatcher(true), match_conf_(match_conf) {} 121 void match(const ImageFeatures &features1, const ImageFeatures &features2, MatchesInfo& matches_info); 122 123 private: 124 float match_conf_; 125 }; 126 127 #ifdef HAVE_OPENCV_CUDAFEATURES2D 128 class GpuMatcher : public FeaturesMatcher 129 { 130 public: 131 GpuMatcher(float match_conf) : match_conf_(match_conf) {} 132 void match(const ImageFeatures &features1, const ImageFeatures &features2, MatchesInfo& matches_info); 133 134 void collectGarbage(); 135 136 private: 137 float match_conf_; 138 GpuMat descriptors1_, descriptors2_; 139 GpuMat train_idx_, distance_, all_dist_; 140 std::vector< std::vector<DMatch> > pair_matches; 141 }; 142 #endif 143 144 145 void CpuMatcher::match(const ImageFeatures &features1, const ImageFeatures &features2, MatchesInfo& matches_info) 146 { 147 CV_Assert(features1.descriptors.type() == features2.descriptors.type()); 148 CV_Assert(features2.descriptors.depth() == CV_8U || features2.descriptors.depth() == CV_32F); 149 150 #ifdef HAVE_TEGRA_OPTIMIZATION 151 if (tegra::useTegra() && tegra::match2nearest(features1, features2, matches_info, match_conf_)) 152 return; 153 #endif 154 155 matches_info.matches.clear(); 156 157 Ptr<cv::DescriptorMatcher> matcher; 158 #if 0 // TODO check this 159 if (ocl::useOpenCL()) 160 { 161 matcher = makePtr<BFMatcher>((int)NORM_L2); 162 } 163 else 164 #endif 165 { 166 Ptr<flann::IndexParams> indexParams = makePtr<flann::KDTreeIndexParams>(); 167 Ptr<flann::SearchParams> searchParams = makePtr<flann::SearchParams>(); 168 169 if (features2.descriptors.depth() == CV_8U) 170 { 171 indexParams->setAlgorithm(cvflann::FLANN_INDEX_LSH); 172 searchParams->setAlgorithm(cvflann::FLANN_INDEX_LSH); 173 } 174 175 matcher = makePtr<FlannBasedMatcher>(indexParams, searchParams); 176 } 177 std::vector< std::vector<DMatch> > pair_matches; 178 MatchesSet matches; 179 180 // Find 1->2 matches 181 matcher->knnMatch(features1.descriptors, features2.descriptors, pair_matches, 2); 182 for (size_t i = 0; i < pair_matches.size(); ++i) 183 { 184 if (pair_matches[i].size() < 2) 185 continue; 186 const DMatch& m0 = pair_matches[i][0]; 187 const DMatch& m1 = pair_matches[i][1]; 188 if (m0.distance < (1.f - match_conf_) * m1.distance) 189 { 190 matches_info.matches.push_back(m0); 191 matches.insert(std::make_pair(m0.queryIdx, m0.trainIdx)); 192 } 193 } 194 LOG("\n1->2 matches: " << matches_info.matches.size() << endl); 195 196 // Find 2->1 matches 197 pair_matches.clear(); 198 matcher->knnMatch(features2.descriptors, features1.descriptors, pair_matches, 2); 199 for (size_t i = 0; i < pair_matches.size(); ++i) 200 { 201 if (pair_matches[i].size() < 2) 202 continue; 203 const DMatch& m0 = pair_matches[i][0]; 204 const DMatch& m1 = pair_matches[i][1]; 205 if (m0.distance < (1.f - match_conf_) * m1.distance) 206 if (matches.find(std::make_pair(m0.trainIdx, m0.queryIdx)) == matches.end()) 207 matches_info.matches.push_back(DMatch(m0.trainIdx, m0.queryIdx, m0.distance)); 208 } 209 LOG("1->2 & 2->1 matches: " << matches_info.matches.size() << endl); 210 } 211 212 #ifdef HAVE_OPENCV_CUDAFEATURES2D 213 void GpuMatcher::match(const ImageFeatures &features1, const ImageFeatures &features2, MatchesInfo& matches_info) 214 { 215 matches_info.matches.clear(); 216 217 ensureSizeIsEnough(features1.descriptors.size(), features1.descriptors.type(), descriptors1_); 218 ensureSizeIsEnough(features2.descriptors.size(), features2.descriptors.type(), descriptors2_); 219 220 descriptors1_.upload(features1.descriptors); 221 descriptors2_.upload(features2.descriptors); 222 223 Ptr<cuda::DescriptorMatcher> matcher = cuda::DescriptorMatcher::createBFMatcher(NORM_L2); 224 225 MatchesSet matches; 226 227 // Find 1->2 matches 228 pair_matches.clear(); 229 matcher->knnMatch(descriptors1_, descriptors2_, pair_matches, 2); 230 for (size_t i = 0; i < pair_matches.size(); ++i) 231 { 232 if (pair_matches[i].size() < 2) 233 continue; 234 const DMatch& m0 = pair_matches[i][0]; 235 const DMatch& m1 = pair_matches[i][1]; 236 if (m0.distance < (1.f - match_conf_) * m1.distance) 237 { 238 matches_info.matches.push_back(m0); 239 matches.insert(std::make_pair(m0.queryIdx, m0.trainIdx)); 240 } 241 } 242 243 // Find 2->1 matches 244 pair_matches.clear(); 245 matcher->knnMatch(descriptors2_, descriptors1_, pair_matches, 2); 246 for (size_t i = 0; i < pair_matches.size(); ++i) 247 { 248 if (pair_matches[i].size() < 2) 249 continue; 250 const DMatch& m0 = pair_matches[i][0]; 251 const DMatch& m1 = pair_matches[i][1]; 252 if (m0.distance < (1.f - match_conf_) * m1.distance) 253 if (matches.find(std::make_pair(m0.trainIdx, m0.queryIdx)) == matches.end()) 254 matches_info.matches.push_back(DMatch(m0.trainIdx, m0.queryIdx, m0.distance)); 255 } 256 } 257 258 void GpuMatcher::collectGarbage() 259 { 260 descriptors1_.release(); 261 descriptors2_.release(); 262 train_idx_.release(); 263 distance_.release(); 264 all_dist_.release(); 265 std::vector< std::vector<DMatch> >().swap(pair_matches); 266 } 267 #endif 268 269 } // namespace 270 271 272 namespace cv { 273 namespace detail { 274 275 void FeaturesFinder::operator ()(InputArray image, ImageFeatures &features) 276 { 277 find(image, features); 278 features.img_size = image.size(); 279 } 280 281 282 void FeaturesFinder::operator ()(InputArray image, ImageFeatures &features, const std::vector<Rect> &rois) 283 { 284 std::vector<ImageFeatures> roi_features(rois.size()); 285 size_t total_kps_count = 0; 286 int total_descriptors_height = 0; 287 288 for (size_t i = 0; i < rois.size(); ++i) 289 { 290 find(image.getUMat()(rois[i]), roi_features[i]); 291 total_kps_count += roi_features[i].keypoints.size(); 292 total_descriptors_height += roi_features[i].descriptors.rows; 293 } 294 295 features.img_size = image.size(); 296 features.keypoints.resize(total_kps_count); 297 features.descriptors.create(total_descriptors_height, 298 roi_features[0].descriptors.cols, 299 roi_features[0].descriptors.type()); 300 301 int kp_idx = 0; 302 int descr_offset = 0; 303 for (size_t i = 0; i < rois.size(); ++i) 304 { 305 for (size_t j = 0; j < roi_features[i].keypoints.size(); ++j, ++kp_idx) 306 { 307 features.keypoints[kp_idx] = roi_features[i].keypoints[j]; 308 features.keypoints[kp_idx].pt.x += (float)rois[i].x; 309 features.keypoints[kp_idx].pt.y += (float)rois[i].y; 310 } 311 UMat subdescr = features.descriptors.rowRange( 312 descr_offset, descr_offset + roi_features[i].descriptors.rows); 313 roi_features[i].descriptors.copyTo(subdescr); 314 descr_offset += roi_features[i].descriptors.rows; 315 } 316 } 317 318 319 SurfFeaturesFinder::SurfFeaturesFinder(double hess_thresh, int num_octaves, int num_layers, 320 int num_octaves_descr, int num_layers_descr) 321 { 322 #ifdef HAVE_OPENCV_XFEATURES2D 323 if (num_octaves_descr == num_octaves && num_layers_descr == num_layers) 324 { 325 Ptr<SURF> surf_ = SURF::create(); 326 if( !surf_ ) 327 CV_Error( Error::StsNotImplemented, "OpenCV was built without SURF support" ); 328 surf_->setHessianThreshold(hess_thresh); 329 surf_->setNOctaves(num_octaves); 330 surf_->setNOctaveLayers(num_layers); 331 surf = surf_; 332 } 333 else 334 { 335 Ptr<SURF> sdetector_ = SURF::create(); 336 Ptr<SURF> sextractor_ = SURF::create(); 337 338 if( !sdetector_ || !sextractor_ ) 339 CV_Error( Error::StsNotImplemented, "OpenCV was built without SURF support" ); 340 341 sdetector_->setHessianThreshold(hess_thresh); 342 sdetector_->setNOctaves(num_octaves); 343 sdetector_->setNOctaveLayers(num_layers); 344 345 sextractor_->setNOctaves(num_octaves_descr); 346 sextractor_->setNOctaveLayers(num_layers_descr); 347 348 detector_ = sdetector_; 349 extractor_ = sextractor_; 350 } 351 #else 352 (void)hess_thresh; 353 (void)num_octaves; 354 (void)num_layers; 355 (void)num_octaves_descr; 356 (void)num_layers_descr; 357 CV_Error( Error::StsNotImplemented, "OpenCV was built without SURF support" ); 358 #endif 359 } 360 361 void SurfFeaturesFinder::find(InputArray image, ImageFeatures &features) 362 { 363 UMat gray_image; 364 CV_Assert((image.type() == CV_8UC3) || (image.type() == CV_8UC1)); 365 if(image.type() == CV_8UC3) 366 { 367 cvtColor(image, gray_image, COLOR_BGR2GRAY); 368 } 369 else 370 { 371 gray_image = image.getUMat(); 372 } 373 if (!surf) 374 { 375 detector_->detect(gray_image, features.keypoints); 376 extractor_->compute(gray_image, features.keypoints, features.descriptors); 377 } 378 else 379 { 380 UMat descriptors; 381 surf->detectAndCompute(gray_image, Mat(), features.keypoints, descriptors); 382 features.descriptors = descriptors.reshape(1, (int)features.keypoints.size()); 383 } 384 } 385 386 OrbFeaturesFinder::OrbFeaturesFinder(Size _grid_size, int n_features, float scaleFactor, int nlevels) 387 { 388 grid_size = _grid_size; 389 orb = ORB::create(n_features * (99 + grid_size.area())/100/grid_size.area(), scaleFactor, nlevels); 390 } 391 392 void OrbFeaturesFinder::find(InputArray image, ImageFeatures &features) 393 { 394 UMat gray_image; 395 396 CV_Assert((image.type() == CV_8UC3) || (image.type() == CV_8UC4) || (image.type() == CV_8UC1)); 397 398 if (image.type() == CV_8UC3) { 399 cvtColor(image, gray_image, COLOR_BGR2GRAY); 400 } else if (image.type() == CV_8UC4) { 401 cvtColor(image, gray_image, COLOR_BGRA2GRAY); 402 } else if (image.type() == CV_8UC1) { 403 gray_image = image.getUMat(); 404 } else { 405 CV_Error(Error::StsUnsupportedFormat, ""); 406 } 407 408 if (grid_size.area() == 1) 409 orb->detectAndCompute(gray_image, Mat(), features.keypoints, features.descriptors); 410 else 411 { 412 features.keypoints.clear(); 413 features.descriptors.release(); 414 415 std::vector<KeyPoint> points; 416 Mat _descriptors; 417 UMat descriptors; 418 419 for (int r = 0; r < grid_size.height; ++r) 420 for (int c = 0; c < grid_size.width; ++c) 421 { 422 int xl = c * gray_image.cols / grid_size.width; 423 int yl = r * gray_image.rows / grid_size.height; 424 int xr = (c+1) * gray_image.cols / grid_size.width; 425 int yr = (r+1) * gray_image.rows / grid_size.height; 426 427 // LOGLN("OrbFeaturesFinder::find: gray_image.empty=" << (gray_image.empty()?"true":"false") << ", " 428 // << " gray_image.size()=(" << gray_image.size().width << "x" << gray_image.size().height << "), " 429 // << " yl=" << yl << ", yr=" << yr << ", " 430 // << " xl=" << xl << ", xr=" << xr << ", gray_image.data=" << ((size_t)gray_image.data) << ", " 431 // << "gray_image.dims=" << gray_image.dims << "\n"); 432 433 UMat gray_image_part=gray_image(Range(yl, yr), Range(xl, xr)); 434 // LOGLN("OrbFeaturesFinder::find: gray_image_part.empty=" << (gray_image_part.empty()?"true":"false") << ", " 435 // << " gray_image_part.size()=(" << gray_image_part.size().width << "x" << gray_image_part.size().height << "), " 436 // << " gray_image_part.dims=" << gray_image_part.dims << ", " 437 // << " gray_image_part.data=" << ((size_t)gray_image_part.data) << "\n"); 438 439 orb->detectAndCompute(gray_image_part, UMat(), points, descriptors); 440 441 features.keypoints.reserve(features.keypoints.size() + points.size()); 442 for (std::vector<KeyPoint>::iterator kp = points.begin(); kp != points.end(); ++kp) 443 { 444 kp->pt.x += xl; 445 kp->pt.y += yl; 446 features.keypoints.push_back(*kp); 447 } 448 _descriptors.push_back(descriptors.getMat(ACCESS_READ)); 449 } 450 451 // TODO optimize copyTo() 452 //features.descriptors = _descriptors.getUMat(ACCESS_READ); 453 _descriptors.copyTo(features.descriptors); 454 } 455 } 456 457 #ifdef HAVE_OPENCV_XFEATURES2D 458 SurfFeaturesFinderGpu::SurfFeaturesFinderGpu(double hess_thresh, int num_octaves, int num_layers, 459 int num_octaves_descr, int num_layers_descr) 460 { 461 surf_.keypointsRatio = 0.1f; 462 surf_.hessianThreshold = hess_thresh; 463 surf_.extended = false; 464 num_octaves_ = num_octaves; 465 num_layers_ = num_layers; 466 num_octaves_descr_ = num_octaves_descr; 467 num_layers_descr_ = num_layers_descr; 468 } 469 470 471 void SurfFeaturesFinderGpu::find(InputArray image, ImageFeatures &features) 472 { 473 CV_Assert(image.depth() == CV_8U); 474 475 ensureSizeIsEnough(image.size(), image.type(), image_); 476 image_.upload(image); 477 478 ensureSizeIsEnough(image.size(), CV_8UC1, gray_image_); 479 cvtColor(image_, gray_image_, COLOR_BGR2GRAY); 480 481 surf_.nOctaves = num_octaves_; 482 surf_.nOctaveLayers = num_layers_; 483 surf_.upright = false; 484 surf_(gray_image_, GpuMat(), keypoints_); 485 486 surf_.nOctaves = num_octaves_descr_; 487 surf_.nOctaveLayers = num_layers_descr_; 488 surf_.upright = true; 489 surf_(gray_image_, GpuMat(), keypoints_, descriptors_, true); 490 surf_.downloadKeypoints(keypoints_, features.keypoints); 491 492 descriptors_.download(features.descriptors); 493 } 494 495 void SurfFeaturesFinderGpu::collectGarbage() 496 { 497 surf_.releaseMemory(); 498 image_.release(); 499 gray_image_.release(); 500 keypoints_.release(); 501 descriptors_.release(); 502 } 503 #endif 504 505 506 ////////////////////////////////////////////////////////////////////////////// 507 508 MatchesInfo::MatchesInfo() : src_img_idx(-1), dst_img_idx(-1), num_inliers(0), confidence(0) {} 509 510 MatchesInfo::MatchesInfo(const MatchesInfo &other) { *this = other; } 511 512 const MatchesInfo& MatchesInfo::operator =(const MatchesInfo &other) 513 { 514 src_img_idx = other.src_img_idx; 515 dst_img_idx = other.dst_img_idx; 516 matches = other.matches; 517 inliers_mask = other.inliers_mask; 518 num_inliers = other.num_inliers; 519 H = other.H.clone(); 520 confidence = other.confidence; 521 return *this; 522 } 523 524 525 ////////////////////////////////////////////////////////////////////////////// 526 527 void FeaturesMatcher::operator ()(const std::vector<ImageFeatures> &features, std::vector<MatchesInfo> &pairwise_matches, 528 const UMat &mask) 529 { 530 const int num_images = static_cast<int>(features.size()); 531 532 CV_Assert(mask.empty() || (mask.type() == CV_8U && mask.cols == num_images && mask.rows)); 533 Mat_<uchar> mask_(mask.getMat(ACCESS_READ)); 534 if (mask_.empty()) 535 mask_ = Mat::ones(num_images, num_images, CV_8U); 536 537 std::vector<std::pair<int,int> > near_pairs; 538 for (int i = 0; i < num_images - 1; ++i) 539 for (int j = i + 1; j < num_images; ++j) 540 if (features[i].keypoints.size() > 0 && features[j].keypoints.size() > 0 && mask_(i, j)) 541 near_pairs.push_back(std::make_pair(i, j)); 542 543 pairwise_matches.resize(num_images * num_images); 544 MatchPairsBody body(*this, features, pairwise_matches, near_pairs); 545 546 if (is_thread_safe_) 547 parallel_for_(Range(0, static_cast<int>(near_pairs.size())), body); 548 else 549 body(Range(0, static_cast<int>(near_pairs.size()))); 550 LOGLN_CHAT(""); 551 } 552 553 554 ////////////////////////////////////////////////////////////////////////////// 555 556 BestOf2NearestMatcher::BestOf2NearestMatcher(bool try_use_gpu, float match_conf, int num_matches_thresh1, int num_matches_thresh2) 557 { 558 (void)try_use_gpu; 559 560 #ifdef HAVE_OPENCV_CUDAFEATURES2D 561 if (try_use_gpu && getCudaEnabledDeviceCount() > 0) 562 { 563 impl_ = makePtr<GpuMatcher>(match_conf); 564 } 565 else 566 #endif 567 { 568 impl_ = makePtr<CpuMatcher>(match_conf); 569 } 570 571 is_thread_safe_ = impl_->isThreadSafe(); 572 num_matches_thresh1_ = num_matches_thresh1; 573 num_matches_thresh2_ = num_matches_thresh2; 574 } 575 576 577 void BestOf2NearestMatcher::match(const ImageFeatures &features1, const ImageFeatures &features2, 578 MatchesInfo &matches_info) 579 { 580 (*impl_)(features1, features2, matches_info); 581 582 // Check if it makes sense to find homography 583 if (matches_info.matches.size() < static_cast<size_t>(num_matches_thresh1_)) 584 return; 585 586 // Construct point-point correspondences for homography estimation 587 Mat src_points(1, static_cast<int>(matches_info.matches.size()), CV_32FC2); 588 Mat dst_points(1, static_cast<int>(matches_info.matches.size()), CV_32FC2); 589 for (size_t i = 0; i < matches_info.matches.size(); ++i) 590 { 591 const DMatch& m = matches_info.matches[i]; 592 593 Point2f p = features1.keypoints[m.queryIdx].pt; 594 p.x -= features1.img_size.width * 0.5f; 595 p.y -= features1.img_size.height * 0.5f; 596 src_points.at<Point2f>(0, static_cast<int>(i)) = p; 597 598 p = features2.keypoints[m.trainIdx].pt; 599 p.x -= features2.img_size.width * 0.5f; 600 p.y -= features2.img_size.height * 0.5f; 601 dst_points.at<Point2f>(0, static_cast<int>(i)) = p; 602 } 603 604 // Find pair-wise motion 605 matches_info.H = findHomography(src_points, dst_points, matches_info.inliers_mask, RANSAC); 606 if (matches_info.H.empty() || std::abs(determinant(matches_info.H)) < std::numeric_limits<double>::epsilon()) 607 return; 608 609 // Find number of inliers 610 matches_info.num_inliers = 0; 611 for (size_t i = 0; i < matches_info.inliers_mask.size(); ++i) 612 if (matches_info.inliers_mask[i]) 613 matches_info.num_inliers++; 614 615 // These coeffs are from paper M. Brown and D. Lowe. "Automatic Panoramic Image Stitching 616 // using Invariant Features" 617 matches_info.confidence = matches_info.num_inliers / (8 + 0.3 * matches_info.matches.size()); 618 619 // Set zero confidence to remove matches between too close images, as they don't provide 620 // additional information anyway. The threshold was set experimentally. 621 matches_info.confidence = matches_info.confidence > 3. ? 0. : matches_info.confidence; 622 623 // Check if we should try to refine motion 624 if (matches_info.num_inliers < num_matches_thresh2_) 625 return; 626 627 // Construct point-point correspondences for inliers only 628 src_points.create(1, matches_info.num_inliers, CV_32FC2); 629 dst_points.create(1, matches_info.num_inliers, CV_32FC2); 630 int inlier_idx = 0; 631 for (size_t i = 0; i < matches_info.matches.size(); ++i) 632 { 633 if (!matches_info.inliers_mask[i]) 634 continue; 635 636 const DMatch& m = matches_info.matches[i]; 637 638 Point2f p = features1.keypoints[m.queryIdx].pt; 639 p.x -= features1.img_size.width * 0.5f; 640 p.y -= features1.img_size.height * 0.5f; 641 src_points.at<Point2f>(0, inlier_idx) = p; 642 643 p = features2.keypoints[m.trainIdx].pt; 644 p.x -= features2.img_size.width * 0.5f; 645 p.y -= features2.img_size.height * 0.5f; 646 dst_points.at<Point2f>(0, inlier_idx) = p; 647 648 inlier_idx++; 649 } 650 651 // Rerun motion estimation on inliers only 652 matches_info.H = findHomography(src_points, dst_points, RANSAC); 653 } 654 655 void BestOf2NearestMatcher::collectGarbage() 656 { 657 impl_->collectGarbage(); 658 } 659 660 661 BestOf2NearestRangeMatcher::BestOf2NearestRangeMatcher(int range_width, bool try_use_gpu, float match_conf, int num_matches_thresh1, int num_matches_thresh2): BestOf2NearestMatcher(try_use_gpu, match_conf, num_matches_thresh1, num_matches_thresh2) 662 { 663 range_width_ = range_width; 664 } 665 666 667 void BestOf2NearestRangeMatcher::operator ()(const std::vector<ImageFeatures> &features, std::vector<MatchesInfo> &pairwise_matches, 668 const UMat &mask) 669 { 670 const int num_images = static_cast<int>(features.size()); 671 672 CV_Assert(mask.empty() || (mask.type() == CV_8U && mask.cols == num_images && mask.rows)); 673 Mat_<uchar> mask_(mask.getMat(ACCESS_READ)); 674 if (mask_.empty()) 675 mask_ = Mat::ones(num_images, num_images, CV_8U); 676 677 std::vector<std::pair<int,int> > near_pairs; 678 for (int i = 0; i < num_images - 1; ++i) 679 for (int j = i + 1; j < std::min(num_images, i + range_width_); ++j) 680 if (features[i].keypoints.size() > 0 && features[j].keypoints.size() > 0 && mask_(i, j)) 681 near_pairs.push_back(std::make_pair(i, j)); 682 683 pairwise_matches.resize(num_images * num_images); 684 MatchPairsBody body(*this, features, pairwise_matches, near_pairs); 685 686 if (is_thread_safe_) 687 parallel_for_(Range(0, static_cast<int>(near_pairs.size())), body); 688 else 689 body(Range(0, static_cast<int>(near_pairs.size()))); 690 LOGLN_CHAT(""); 691 } 692 693 694 } // namespace detail 695 } // namespace cv 696