1 // Copyright (c) 2013 libmv authors. 2 // 3 // Permission is hereby granted, free of charge, to any person obtaining a copy 4 // of this software and associated documentation files (the "Software"), to 5 // deal in the Software without restriction, including without limitation the 6 // rights to use, copy, modify, merge, publish, distribute, sublicense, and/or 7 // sell copies of the Software, and to permit persons to whom the Software is 8 // furnished to do so, subject to the following conditions: 9 // 10 // The above copyright notice and this permission notice shall be included in 11 // all copies or substantial portions of the Software. 12 // 13 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 16 // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 17 // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 18 // FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 19 // IN THE SOFTWARE. 20 // 21 // Author: mierle (at) gmail.com (Keir Mierle) 22 // sergey.vfx (at) gmail.com (Sergey Sharybin) 23 // 24 // This is an example application which contains bundle adjustment code used 25 // in the Libmv library and Blender. It reads problems from files passed via 26 // the command line and runs the bundle adjuster on the problem. 27 // 28 // File with problem a binary file, for which it is crucial to know in which 29 // order bytes of float values are stored in. This information is provided 30 // by a single character in the beginning of the file. There're two possible 31 // values of this byte: 32 // - V, which means values in the file are stored with big endian type 33 // - v, which means values in the file are stored with little endian type 34 // 35 // The rest of the file contains data in the following order: 36 // - Space in which markers' coordinates are stored in 37 // - Camera intrinsics 38 // - Number of cameras 39 // - Cameras 40 // - Number of 3D points 41 // - 3D points 42 // - Number of markers 43 // - Markers 44 // 45 // Markers' space could either be normalized or image (pixels). This is defined 46 // by the single character in the file. P means markers in the file is in image 47 // space, and N means markers are in normalized space. 48 // 49 // Camera intrinsics are 8 described by 8 float 8. 50 // This values goes in the following order: 51 // 52 // - Focal length, principal point X, principal point Y, k1, k2, k3, p1, p2 53 // 54 // Every camera is described by: 55 // 56 // - Image for which camera belongs to (single 4 bytes integer value). 57 // - Column-major camera rotation matrix, 9 float values. 58 // - Camera translation, 3-component vector of float values. 59 // 60 // Image number shall be greater or equal to zero. Order of cameras does not 61 // matter and gaps are possible. 62 // 63 // Every 3D point is decribed by: 64 // 65 // - Track number point belongs to (single 4 bytes integer value). 66 // - 3D position vector, 3-component vector of float values. 67 // 68 // Track number shall be greater or equal to zero. Order of tracks does not 69 // matter and gaps are possible. 70 // 71 // Finally every marker is described by: 72 // 73 // - Image marker belongs to single 4 bytes integer value). 74 // - Track marker belongs to single 4 bytes integer value). 75 // - 2D marker position vector, (two float values). 76 // 77 // Marker's space is used a default value for refine_intrinsics command line 78 // flag. This means if there's no refine_intrinsics flag passed via command 79 // line, camera intrinsics will be refined if markers in the problem are 80 // stored in image space and camera intrinsics will not be refined if markers 81 // are in normalized space. 82 // 83 // Passing refine_intrinsics command line flag defines explicitly whether 84 // refinement of intrinsics will happen. Currently, only none and all 85 // intrinsics refinement is supported. 86 // 87 // There're existing problem files dumped from blender stored in folder 88 // ../data/libmv-ba-problems. 89 90 #include <cstdio> 91 #include <fcntl.h> 92 #include <sstream> 93 #include <string> 94 #include <vector> 95 96 #ifdef _MSC_VER 97 # include <io.h> 98 # define open _open 99 # define close _close 100 typedef unsigned __int32 uint32_t; 101 #else 102 # include <stdint.h> 103 104 // O_BINARY is not defined on unix like platforms, as there is no 105 // difference between binary and text files. 106 #define O_BINARY 0 107 108 #endif 109 110 #include "ceres/ceres.h" 111 #include "ceres/rotation.h" 112 #include "gflags/gflags.h" 113 #include "glog/logging.h" 114 115 typedef Eigen::Matrix<double, 3, 3> Mat3; 116 typedef Eigen::Matrix<double, 6, 1> Vec6; 117 typedef Eigen::Vector3d Vec3; 118 typedef Eigen::Vector4d Vec4; 119 120 using std::vector; 121 122 DEFINE_string(input, "", "Input File name"); 123 DEFINE_string(refine_intrinsics, "", "Camera intrinsics to be refined. " 124 "Options are: none, radial."); 125 126 namespace { 127 128 // A EuclideanCamera is the location and rotation of the camera 129 // viewing an image. 130 // 131 // image identifies which image this camera represents. 132 // R is a 3x3 matrix representing the rotation of the camera. 133 // t is a translation vector representing its positions. 134 struct EuclideanCamera { 135 EuclideanCamera() : image(-1) {} 136 EuclideanCamera(const EuclideanCamera &c) : image(c.image), R(c.R), t(c.t) {} 137 138 int image; 139 Mat3 R; 140 Vec3 t; 141 }; 142 143 // A Point is the 3D location of a track. 144 // 145 // track identifies which track this point corresponds to. 146 // X represents the 3D position of the track. 147 struct EuclideanPoint { 148 EuclideanPoint() : track(-1) {} 149 EuclideanPoint(const EuclideanPoint &p) : track(p.track), X(p.X) {} 150 int track; 151 Vec3 X; 152 }; 153 154 // A Marker is the 2D location of a tracked point in an image. 155 // 156 // x and y is the position of the marker in pixels from the top left corner 157 // in the image identified by an image. All markers for to the same target 158 // form a track identified by a common track number. 159 struct Marker { 160 int image; 161 int track; 162 double x, y; 163 }; 164 165 // Cameras intrinsics to be bundled. 166 // 167 // BUNDLE_RADIAL actually implies bundling of k1 and k2 coefficients only, 168 // no bundling of k3 is possible at this moment. 169 enum BundleIntrinsics { 170 BUNDLE_NO_INTRINSICS = 0, 171 BUNDLE_FOCAL_LENGTH = 1, 172 BUNDLE_PRINCIPAL_POINT = 2, 173 BUNDLE_RADIAL_K1 = 4, 174 BUNDLE_RADIAL_K2 = 8, 175 BUNDLE_RADIAL = 12, 176 BUNDLE_TANGENTIAL_P1 = 16, 177 BUNDLE_TANGENTIAL_P2 = 32, 178 BUNDLE_TANGENTIAL = 48, 179 }; 180 181 // Denotes which blocks to keep constant during bundling. 182 // For example it is useful to keep camera translations constant 183 // when bundling tripod motions. 184 enum BundleConstraints { 185 BUNDLE_NO_CONSTRAINTS = 0, 186 BUNDLE_NO_TRANSLATION = 1, 187 }; 188 189 // The intrinsics need to get combined into a single parameter block; use these 190 // enums to index instead of numeric constants. 191 enum { 192 OFFSET_FOCAL_LENGTH, 193 OFFSET_PRINCIPAL_POINT_X, 194 OFFSET_PRINCIPAL_POINT_Y, 195 OFFSET_K1, 196 OFFSET_K2, 197 OFFSET_K3, 198 OFFSET_P1, 199 OFFSET_P2, 200 }; 201 202 // Returns a pointer to the camera corresponding to a image. 203 EuclideanCamera *CameraForImage(vector<EuclideanCamera> *all_cameras, 204 const int image) { 205 if (image < 0 || image >= all_cameras->size()) { 206 return NULL; 207 } 208 EuclideanCamera *camera = &(*all_cameras)[image]; 209 if (camera->image == -1) { 210 return NULL; 211 } 212 return camera; 213 } 214 215 const EuclideanCamera *CameraForImage( 216 const vector<EuclideanCamera> &all_cameras, 217 const int image) { 218 if (image < 0 || image >= all_cameras.size()) { 219 return NULL; 220 } 221 const EuclideanCamera *camera = &all_cameras[image]; 222 if (camera->image == -1) { 223 return NULL; 224 } 225 return camera; 226 } 227 228 // Returns maximal image number at which marker exists. 229 int MaxImage(const vector<Marker> &all_markers) { 230 if (all_markers.size() == 0) { 231 return -1; 232 } 233 234 int max_image = all_markers[0].image; 235 for (int i = 1; i < all_markers.size(); i++) { 236 max_image = std::max(max_image, all_markers[i].image); 237 } 238 return max_image; 239 } 240 241 // Returns a pointer to the point corresponding to a track. 242 EuclideanPoint *PointForTrack(vector<EuclideanPoint> *all_points, 243 const int track) { 244 if (track < 0 || track >= all_points->size()) { 245 return NULL; 246 } 247 EuclideanPoint *point = &(*all_points)[track]; 248 if (point->track == -1) { 249 return NULL; 250 } 251 return point; 252 } 253 254 // Reader of binary file which makes sure possibly needed endian 255 // conversion happens when loading values like floats and integers. 256 // 257 // File's endian type is reading from a first character of file, which 258 // could either be V for big endian or v for little endian. This 259 // means you need to design file format assuming first character 260 // denotes file endianness in this way. 261 class EndianAwareFileReader { 262 public: 263 EndianAwareFileReader(void) : file_descriptor_(-1) { 264 // Get an endian type of the host machine. 265 union { 266 unsigned char bytes[4]; 267 uint32_t value; 268 } endian_test = { { 0, 1, 2, 3 } }; 269 host_endian_type_ = endian_test.value; 270 file_endian_type_ = host_endian_type_; 271 } 272 273 ~EndianAwareFileReader(void) { 274 if (file_descriptor_ > 0) { 275 close(file_descriptor_); 276 } 277 } 278 279 bool OpenFile(const std::string &file_name) { 280 file_descriptor_ = open(file_name.c_str(), O_RDONLY | O_BINARY); 281 if (file_descriptor_ < 0) { 282 return false; 283 } 284 // Get an endian tpye of data in the file. 285 unsigned char file_endian_type_flag = Read<unsigned char>(); 286 if (file_endian_type_flag == 'V') { 287 file_endian_type_ = kBigEndian; 288 } else if (file_endian_type_flag == 'v') { 289 file_endian_type_ = kLittleEndian; 290 } else { 291 LOG(FATAL) << "Problem file is stored in unknown endian type."; 292 } 293 return true; 294 } 295 296 // Read value from the file, will switch endian if needed. 297 template <typename T> 298 T Read(void) const { 299 T value; 300 CHECK_GT(read(file_descriptor_, &value, sizeof(value)), 0); 301 // Switch endian type if file contains data in different type 302 // that current machine. 303 if (file_endian_type_ != host_endian_type_) { 304 value = SwitchEndian<T>(value); 305 } 306 return value; 307 } 308 private: 309 static const long int kLittleEndian = 0x03020100ul; 310 static const long int kBigEndian = 0x00010203ul; 311 312 // Switch endian type between big to little. 313 template <typename T> 314 T SwitchEndian(const T value) const { 315 if (sizeof(T) == 4) { 316 unsigned int temp_value = static_cast<unsigned int>(value); 317 return ((temp_value >> 24)) | 318 ((temp_value << 8) & 0x00ff0000) | 319 ((temp_value >> 8) & 0x0000ff00) | 320 ((temp_value << 24)); 321 } else if (sizeof(T) == 1) { 322 return value; 323 } else { 324 LOG(FATAL) << "Entered non-implemented part of endian switching function."; 325 } 326 } 327 328 int host_endian_type_; 329 int file_endian_type_; 330 int file_descriptor_; 331 }; 332 333 // Read 3x3 column-major matrix from the file 334 void ReadMatrix3x3(const EndianAwareFileReader &file_reader, 335 Mat3 *matrix) { 336 for (int i = 0; i < 9; i++) { 337 (*matrix)(i % 3, i / 3) = file_reader.Read<float>(); 338 } 339 } 340 341 // Read 3-vector from file 342 void ReadVector3(const EndianAwareFileReader &file_reader, 343 Vec3 *vector) { 344 for (int i = 0; i < 3; i++) { 345 (*vector)(i) = file_reader.Read<float>(); 346 } 347 } 348 349 // Reads a bundle adjustment problem from the file. 350 // 351 // file_name denotes from which file to read the problem. 352 // camera_intrinsics will contain initial camera intrinsics values. 353 // 354 // all_cameras is a vector of all reconstructed cameras to be optimized, 355 // vector element with number i will contain camera for image i. 356 // 357 // all_points is a vector of all reconstructed 3D points to be optimized, 358 // vector element with number i will contain point for track i. 359 // 360 // all_markers is a vector of all tracked markers existing in 361 // the problem. Only used for reprojection error calculation, stay 362 // unchanged during optimization. 363 // 364 // Returns false if any kind of error happened during 365 // reading. 366 bool ReadProblemFromFile(const std::string &file_name, 367 double camera_intrinsics[8], 368 vector<EuclideanCamera> *all_cameras, 369 vector<EuclideanPoint> *all_points, 370 bool *is_image_space, 371 vector<Marker> *all_markers) { 372 EndianAwareFileReader file_reader; 373 if (!file_reader.OpenFile(file_name)) { 374 return false; 375 } 376 377 // Read markers' space flag. 378 unsigned char is_image_space_flag = file_reader.Read<unsigned char>(); 379 if (is_image_space_flag == 'P') { 380 *is_image_space = true; 381 } else if (is_image_space_flag == 'N') { 382 *is_image_space = false; 383 } else { 384 LOG(FATAL) << "Problem file contains markers stored in unknown space."; 385 } 386 387 // Read camera intrinsics. 388 for (int i = 0; i < 8; i++) { 389 camera_intrinsics[i] = file_reader.Read<float>(); 390 } 391 392 // Read all cameras. 393 int number_of_cameras = file_reader.Read<int>(); 394 for (int i = 0; i < number_of_cameras; i++) { 395 EuclideanCamera camera; 396 397 camera.image = file_reader.Read<int>(); 398 ReadMatrix3x3(file_reader, &camera.R); 399 ReadVector3(file_reader, &camera.t); 400 401 if (camera.image >= all_cameras->size()) { 402 all_cameras->resize(camera.image + 1); 403 } 404 405 (*all_cameras)[camera.image].image = camera.image; 406 (*all_cameras)[camera.image].R = camera.R; 407 (*all_cameras)[camera.image].t = camera.t; 408 } 409 410 LOG(INFO) << "Read " << number_of_cameras << " cameras."; 411 412 // Read all reconstructed 3D points. 413 int number_of_points = file_reader.Read<int>(); 414 for (int i = 0; i < number_of_points; i++) { 415 EuclideanPoint point; 416 417 point.track = file_reader.Read<int>(); 418 ReadVector3(file_reader, &point.X); 419 420 if (point.track >= all_points->size()) { 421 all_points->resize(point.track + 1); 422 } 423 424 (*all_points)[point.track].track = point.track; 425 (*all_points)[point.track].X = point.X; 426 } 427 428 LOG(INFO) << "Read " << number_of_points << " points."; 429 430 // And finally read all markers. 431 int number_of_markers = file_reader.Read<int>(); 432 for (int i = 0; i < number_of_markers; i++) { 433 Marker marker; 434 435 marker.image = file_reader.Read<int>(); 436 marker.track = file_reader.Read<int>(); 437 marker.x = file_reader.Read<float>(); 438 marker.y = file_reader.Read<float>(); 439 440 all_markers->push_back(marker); 441 } 442 443 LOG(INFO) << "Read " << number_of_markers << " markers."; 444 445 return true; 446 } 447 448 // Apply camera intrinsics to the normalized point to get image coordinates. 449 // This applies the radial lens distortion to a point which is in normalized 450 // camera coordinates (i.e. the principal point is at (0, 0)) to get image 451 // coordinates in pixels. Templated for use with autodifferentiation. 452 template <typename T> 453 inline void ApplyRadialDistortionCameraIntrinsics(const T &focal_length_x, 454 const T &focal_length_y, 455 const T &principal_point_x, 456 const T &principal_point_y, 457 const T &k1, 458 const T &k2, 459 const T &k3, 460 const T &p1, 461 const T &p2, 462 const T &normalized_x, 463 const T &normalized_y, 464 T *image_x, 465 T *image_y) { 466 T x = normalized_x; 467 T y = normalized_y; 468 469 // Apply distortion to the normalized points to get (xd, yd). 470 T r2 = x*x + y*y; 471 T r4 = r2 * r2; 472 T r6 = r4 * r2; 473 T r_coeff = (T(1) + k1*r2 + k2*r4 + k3*r6); 474 T xd = x * r_coeff + T(2)*p1*x*y + p2*(r2 + T(2)*x*x); 475 T yd = y * r_coeff + T(2)*p2*x*y + p1*(r2 + T(2)*y*y); 476 477 // Apply focal length and principal point to get the final image coordinates. 478 *image_x = focal_length_x * xd + principal_point_x; 479 *image_y = focal_length_y * yd + principal_point_y; 480 } 481 482 // Cost functor which computes reprojection error of 3D point X 483 // on camera defined by angle-axis rotation and it's translation 484 // (which are in the same block due to optimization reasons). 485 // 486 // This functor uses a radial distortion model. 487 struct OpenCVReprojectionError { 488 OpenCVReprojectionError(const double observed_x, const double observed_y) 489 : observed_x(observed_x), observed_y(observed_y) {} 490 491 template <typename T> 492 bool operator()(const T* const intrinsics, 493 const T* const R_t, // Rotation denoted by angle axis 494 // followed with translation 495 const T* const X, // Point coordinates 3x1. 496 T* residuals) const { 497 // Unpack the intrinsics. 498 const T& focal_length = intrinsics[OFFSET_FOCAL_LENGTH]; 499 const T& principal_point_x = intrinsics[OFFSET_PRINCIPAL_POINT_X]; 500 const T& principal_point_y = intrinsics[OFFSET_PRINCIPAL_POINT_Y]; 501 const T& k1 = intrinsics[OFFSET_K1]; 502 const T& k2 = intrinsics[OFFSET_K2]; 503 const T& k3 = intrinsics[OFFSET_K3]; 504 const T& p1 = intrinsics[OFFSET_P1]; 505 const T& p2 = intrinsics[OFFSET_P2]; 506 507 // Compute projective coordinates: x = RX + t. 508 T x[3]; 509 510 ceres::AngleAxisRotatePoint(R_t, X, x); 511 x[0] += R_t[3]; 512 x[1] += R_t[4]; 513 x[2] += R_t[5]; 514 515 // Compute normalized coordinates: x /= x[2]. 516 T xn = x[0] / x[2]; 517 T yn = x[1] / x[2]; 518 519 T predicted_x, predicted_y; 520 521 // Apply distortion to the normalized points to get (xd, yd). 522 // TODO(keir): Do early bailouts for zero distortion; these are expensive 523 // jet operations. 524 ApplyRadialDistortionCameraIntrinsics(focal_length, 525 focal_length, 526 principal_point_x, 527 principal_point_y, 528 k1, k2, k3, 529 p1, p2, 530 xn, yn, 531 &predicted_x, 532 &predicted_y); 533 534 // The error is the difference between the predicted and observed position. 535 residuals[0] = predicted_x - T(observed_x); 536 residuals[1] = predicted_y - T(observed_y); 537 538 return true; 539 } 540 541 const double observed_x; 542 const double observed_y; 543 }; 544 545 // Print a message to the log which camera intrinsics are gonna to be optimized. 546 void BundleIntrinsicsLogMessage(const int bundle_intrinsics) { 547 if (bundle_intrinsics == BUNDLE_NO_INTRINSICS) { 548 LOG(INFO) << "Bundling only camera positions."; 549 } else { 550 std::string bundling_message = ""; 551 552 #define APPEND_BUNDLING_INTRINSICS(name, flag) \ 553 if (bundle_intrinsics & flag) { \ 554 if (!bundling_message.empty()) { \ 555 bundling_message += ", "; \ 556 } \ 557 bundling_message += name; \ 558 } (void)0 559 560 APPEND_BUNDLING_INTRINSICS("f", BUNDLE_FOCAL_LENGTH); 561 APPEND_BUNDLING_INTRINSICS("px, py", BUNDLE_PRINCIPAL_POINT); 562 APPEND_BUNDLING_INTRINSICS("k1", BUNDLE_RADIAL_K1); 563 APPEND_BUNDLING_INTRINSICS("k2", BUNDLE_RADIAL_K2); 564 APPEND_BUNDLING_INTRINSICS("p1", BUNDLE_TANGENTIAL_P1); 565 APPEND_BUNDLING_INTRINSICS("p2", BUNDLE_TANGENTIAL_P2); 566 567 LOG(INFO) << "Bundling " << bundling_message << "."; 568 } 569 } 570 571 // Print a message to the log containing all the camera intriniscs values. 572 void PrintCameraIntrinsics(const char *text, const double *camera_intrinsics) { 573 std::ostringstream intrinsics_output; 574 575 intrinsics_output << "f=" << camera_intrinsics[OFFSET_FOCAL_LENGTH]; 576 577 intrinsics_output << 578 " cx=" << camera_intrinsics[OFFSET_PRINCIPAL_POINT_X] << 579 " cy=" << camera_intrinsics[OFFSET_PRINCIPAL_POINT_Y]; 580 581 #define APPEND_DISTORTION_COEFFICIENT(name, offset) \ 582 { \ 583 if (camera_intrinsics[offset] != 0.0) { \ 584 intrinsics_output << " " name "=" << camera_intrinsics[offset]; \ 585 } \ 586 } (void)0 587 588 APPEND_DISTORTION_COEFFICIENT("k1", OFFSET_K1); 589 APPEND_DISTORTION_COEFFICIENT("k2", OFFSET_K2); 590 APPEND_DISTORTION_COEFFICIENT("k3", OFFSET_K3); 591 APPEND_DISTORTION_COEFFICIENT("p1", OFFSET_P1); 592 APPEND_DISTORTION_COEFFICIENT("p2", OFFSET_P2); 593 594 #undef APPEND_DISTORTION_COEFFICIENT 595 596 LOG(INFO) << text << intrinsics_output.str(); 597 } 598 599 // Get a vector of camera's rotations denoted by angle axis 600 // conjuncted with translations into single block 601 // 602 // Element with index i matches to a rotation+translation for 603 // camera at image i. 604 vector<Vec6> PackCamerasRotationAndTranslation( 605 const vector<Marker> &all_markers, 606 const vector<EuclideanCamera> &all_cameras) { 607 vector<Vec6> all_cameras_R_t; 608 int max_image = MaxImage(all_markers); 609 610 all_cameras_R_t.resize(max_image + 1); 611 612 for (int i = 0; i <= max_image; i++) { 613 const EuclideanCamera *camera = CameraForImage(all_cameras, i); 614 615 if (!camera) { 616 continue; 617 } 618 619 ceres::RotationMatrixToAngleAxis(&camera->R(0, 0), 620 &all_cameras_R_t[i](0)); 621 all_cameras_R_t[i].tail<3>() = camera->t; 622 } 623 624 return all_cameras_R_t; 625 } 626 627 // Convert cameras rotations fro mangle axis back to rotation matrix. 628 void UnpackCamerasRotationAndTranslation( 629 const vector<Marker> &all_markers, 630 const vector<Vec6> &all_cameras_R_t, 631 vector<EuclideanCamera> *all_cameras) { 632 int max_image = MaxImage(all_markers); 633 634 for (int i = 0; i <= max_image; i++) { 635 EuclideanCamera *camera = CameraForImage(all_cameras, i); 636 637 if (!camera) { 638 continue; 639 } 640 641 ceres::AngleAxisToRotationMatrix(&all_cameras_R_t[i](0), 642 &camera->R(0, 0)); 643 camera->t = all_cameras_R_t[i].tail<3>(); 644 } 645 } 646 647 void EuclideanBundleCommonIntrinsics(const vector<Marker> &all_markers, 648 const int bundle_intrinsics, 649 const int bundle_constraints, 650 double *camera_intrinsics, 651 vector<EuclideanCamera> *all_cameras, 652 vector<EuclideanPoint> *all_points) { 653 PrintCameraIntrinsics("Original intrinsics: ", camera_intrinsics); 654 655 ceres::Problem::Options problem_options; 656 ceres::Problem problem(problem_options); 657 658 // Convert cameras rotations to angle axis and merge with translation 659 // into single parameter block for maximal minimization speed 660 // 661 // Block for minimization has got the following structure: 662 // <3 elements for angle-axis> <3 elements for translation> 663 vector<Vec6> all_cameras_R_t = 664 PackCamerasRotationAndTranslation(all_markers, *all_cameras); 665 666 // Parameterization used to restrict camera motion for modal solvers. 667 ceres::SubsetParameterization *constant_transform_parameterization = NULL; 668 if (bundle_constraints & BUNDLE_NO_TRANSLATION) { 669 std::vector<int> constant_translation; 670 671 // First three elements are rotation, last three are translation. 672 constant_translation.push_back(3); 673 constant_translation.push_back(4); 674 constant_translation.push_back(5); 675 676 constant_transform_parameterization = 677 new ceres::SubsetParameterization(6, constant_translation); 678 } 679 680 int num_residuals = 0; 681 bool have_locked_camera = false; 682 for (int i = 0; i < all_markers.size(); ++i) { 683 const Marker &marker = all_markers[i]; 684 EuclideanCamera *camera = CameraForImage(all_cameras, marker.image); 685 EuclideanPoint *point = PointForTrack(all_points, marker.track); 686 if (camera == NULL || point == NULL) { 687 continue; 688 } 689 690 // Rotation of camera denoted in angle axis followed with 691 // camera translaiton. 692 double *current_camera_R_t = &all_cameras_R_t[camera->image](0); 693 694 problem.AddResidualBlock(new ceres::AutoDiffCostFunction< 695 OpenCVReprojectionError, 2, 8, 6, 3>( 696 new OpenCVReprojectionError( 697 marker.x, 698 marker.y)), 699 NULL, 700 camera_intrinsics, 701 current_camera_R_t, 702 &point->X(0)); 703 704 // We lock the first camera to better deal with scene orientation ambiguity. 705 if (!have_locked_camera) { 706 problem.SetParameterBlockConstant(current_camera_R_t); 707 have_locked_camera = true; 708 } 709 710 if (bundle_constraints & BUNDLE_NO_TRANSLATION) { 711 problem.SetParameterization(current_camera_R_t, 712 constant_transform_parameterization); 713 } 714 715 num_residuals++; 716 } 717 LOG(INFO) << "Number of residuals: " << num_residuals; 718 719 if (!num_residuals) { 720 LOG(INFO) << "Skipping running minimizer with zero residuals"; 721 return; 722 } 723 724 BundleIntrinsicsLogMessage(bundle_intrinsics); 725 726 if (bundle_intrinsics == BUNDLE_NO_INTRINSICS) { 727 // No camera intrinsics are being refined, 728 // set the whole parameter block as constant for best performance. 729 problem.SetParameterBlockConstant(camera_intrinsics); 730 } else { 731 // Set the camera intrinsics that are not to be bundled as 732 // constant using some macro trickery. 733 734 std::vector<int> constant_intrinsics; 735 #define MAYBE_SET_CONSTANT(bundle_enum, offset) \ 736 if (!(bundle_intrinsics & bundle_enum)) { \ 737 constant_intrinsics.push_back(offset); \ 738 } 739 MAYBE_SET_CONSTANT(BUNDLE_FOCAL_LENGTH, OFFSET_FOCAL_LENGTH); 740 MAYBE_SET_CONSTANT(BUNDLE_PRINCIPAL_POINT, OFFSET_PRINCIPAL_POINT_X); 741 MAYBE_SET_CONSTANT(BUNDLE_PRINCIPAL_POINT, OFFSET_PRINCIPAL_POINT_Y); 742 MAYBE_SET_CONSTANT(BUNDLE_RADIAL_K1, OFFSET_K1); 743 MAYBE_SET_CONSTANT(BUNDLE_RADIAL_K2, OFFSET_K2); 744 MAYBE_SET_CONSTANT(BUNDLE_TANGENTIAL_P1, OFFSET_P1); 745 MAYBE_SET_CONSTANT(BUNDLE_TANGENTIAL_P2, OFFSET_P2); 746 #undef MAYBE_SET_CONSTANT 747 748 // Always set K3 constant, it's not used at the moment. 749 constant_intrinsics.push_back(OFFSET_K3); 750 751 ceres::SubsetParameterization *subset_parameterization = 752 new ceres::SubsetParameterization(8, constant_intrinsics); 753 754 problem.SetParameterization(camera_intrinsics, subset_parameterization); 755 } 756 757 // Configure the solver. 758 ceres::Solver::Options options; 759 options.use_nonmonotonic_steps = true; 760 options.preconditioner_type = ceres::SCHUR_JACOBI; 761 options.linear_solver_type = ceres::ITERATIVE_SCHUR; 762 options.use_inner_iterations = true; 763 options.max_num_iterations = 100; 764 options.minimizer_progress_to_stdout = true; 765 766 // Solve! 767 ceres::Solver::Summary summary; 768 ceres::Solve(options, &problem, &summary); 769 770 std::cout << "Final report:\n" << summary.FullReport(); 771 772 // Copy rotations and translations back. 773 UnpackCamerasRotationAndTranslation(all_markers, 774 all_cameras_R_t, 775 all_cameras); 776 777 PrintCameraIntrinsics("Final intrinsics: ", camera_intrinsics); 778 } 779 } // namespace 780 781 int main(int argc, char **argv) { 782 google::ParseCommandLineFlags(&argc, &argv, true); 783 google::InitGoogleLogging(argv[0]); 784 785 if (FLAGS_input.empty()) { 786 LOG(ERROR) << "Usage: libmv_bundle_adjuster --input=blender_problem"; 787 return EXIT_FAILURE; 788 } 789 790 double camera_intrinsics[8]; 791 vector<EuclideanCamera> all_cameras; 792 vector<EuclideanPoint> all_points; 793 bool is_image_space; 794 vector<Marker> all_markers; 795 796 if (!ReadProblemFromFile(FLAGS_input, 797 camera_intrinsics, 798 &all_cameras, 799 &all_points, 800 &is_image_space, 801 &all_markers)) { 802 LOG(ERROR) << "Error reading problem file"; 803 return EXIT_FAILURE; 804 } 805 806 // If there's no refine_intrinsics passed via command line 807 // (in this case FLAGS_refine_intrinsics will be an empty string) 808 // we use problem's settings to detect whether intrinsics 809 // shall be refined or not. 810 // 811 // Namely, if problem has got markers stored in image (pixel) 812 // space, we do full intrinsics refinement. If markers are 813 // stored in normalized space, and refine_intrinsics is not 814 // set, no refining will happen. 815 // 816 // Using command line argument refine_intrinsics will explicitly 817 // declare which intrinsics need to be refined and in this case 818 // refining flags does not depend on problem at all. 819 int bundle_intrinsics = BUNDLE_NO_INTRINSICS; 820 if (FLAGS_refine_intrinsics.empty()) { 821 if (is_image_space) { 822 bundle_intrinsics = BUNDLE_FOCAL_LENGTH | BUNDLE_RADIAL; 823 } 824 } else { 825 if (FLAGS_refine_intrinsics == "radial") { 826 bundle_intrinsics = BUNDLE_FOCAL_LENGTH | BUNDLE_RADIAL; 827 } else if (FLAGS_refine_intrinsics != "none") { 828 LOG(ERROR) << "Unsupported value for refine-intrinsics"; 829 return EXIT_FAILURE; 830 } 831 } 832 833 // Run the bundler. 834 EuclideanBundleCommonIntrinsics(all_markers, 835 bundle_intrinsics, 836 BUNDLE_NO_CONSTRAINTS, 837 camera_intrinsics, 838 &all_cameras, 839 &all_points); 840 841 return EXIT_SUCCESS; 842 } 843