opencv/modules/3d/test/test_usac.cpp
Alexander Alekhin fc628014bb Merge branch 4.x
2021-04-10 18:03:01 +00:00

444 lines
20 KiB
C++

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include "test_precomp.hpp"
namespace opencv_test { namespace {
enum TestSolver { Homogr, Fundam, Essen, PnP, Affine};
/*
* rng -- reference to random generator
* pts1 -- 2xN image points
* pts2 -- for PnP is 3xN object points, otherwise 2xN image points.
* two_calib -- True if two cameras have different calibration.
* K1 -- intrinsic matrix of the first camera. For PnP only one camera.
* K2 -- only if two_calib is True.
* pts_size -- required size of points.
* inlier_ratio -- required inlier ratio
* noise_std -- standard deviation of Gaussian noise of image points.
* gt_inliers -- has size of number of inliers. Contains indices of inliers.
*/
static int generatePoints (cv::RNG &rng, cv::Mat &pts1, cv::Mat &pts2, cv::Mat &K1, cv::Mat &K2,
bool two_calib, int pts_size, TestSolver test_case, double inlier_ratio, double noise_std,
std::vector<int> &gt_inliers) {
auto eulerAnglesToRotationMatrix = [] (double pitch, double yaw, double roll) {
// Calculate rotation about x axis
cv::Matx33d R_x (1, 0, 0, 0, cos(roll), -sin(roll), 0, sin(roll), cos(roll));
// Calculate rotation about y axis
cv::Matx33d R_y (cos(pitch), 0, sin(pitch), 0, 1, 0, -sin(pitch), 0, cos(pitch));
// Calculate rotation about z axis
cv::Matx33d R_z (cos(yaw), -sin(yaw), 0, sin(yaw), cos(yaw), 0, 0, 0, 1);
return cv::Mat(R_z * R_y * R_x); // Combined rotation matrix
};
const double pitch_min = -CV_PI / 6, pitch_max = CV_PI / 6; // 30 degrees
const double yaw_min = -CV_PI / 6, yaw_max = CV_PI / 6;
const double roll_min = -CV_PI / 6, roll_max = CV_PI / 6;
cv::Mat R = eulerAnglesToRotationMatrix(rng.uniform(pitch_min, pitch_max),
rng.uniform(yaw_min, yaw_max), rng.uniform(roll_min, roll_max));
// generate random translation,
// if test for homography fails try to fix translation to zero vec so H is related by transl.
cv::Vec3d t (rng.uniform(-0.5f, 0.5f), rng.uniform(-0.5f, 0.5f), rng.uniform(1.0f, 2.0f));
// generate random calibration
auto getRandomCalib = [&] () {
return cv::Mat(cv::Matx33d(rng.uniform(100.0, 1000.0), 0, rng.uniform(100.0, 100.0),
0, rng.uniform(100.0, 1000.0), rng.uniform(-100.0, 100.0),
0, 0, 1.));
};
K1 = getRandomCalib();
K2 = two_calib ? getRandomCalib() : K1.clone();
auto updateTranslation = [] (const cv::Mat &pts, const cv::Mat &R_, cv::Vec3d &t_) {
// Make sure the shape is in front of the camera
cv::Mat points3d_transformed = R_ * pts + t_ * cv::Mat::ones(1, pts.cols, pts.type());
double min_dist, max_dist;
cv::minMaxIdx(points3d_transformed.row(2), &min_dist, &max_dist);
if (min_dist < 0) t_(2) -= min_dist + 1.0;
};
// compute size of inliers and outliers
const int inl_size = static_cast<int>(inlier_ratio * pts_size);
const int out_size = pts_size - inl_size;
// all points will have top 'inl_size' of their points inliers
gt_inliers.clear(); gt_inliers.reserve(inl_size);
for (int i = 0; i < inl_size; i++)
gt_inliers.emplace_back(i);
// double precision to multiply points by models
const int pts_type = CV_64F;
cv::Mat points3d;
if (test_case == TestSolver::Homogr) {
points3d.create(2, inl_size, pts_type);
rng.fill(points3d, cv::RNG::UNIFORM, 0.0, 1.0); // keep small range
// inliers must be planar points, let their 3D coordinate be 1
cv::vconcat(points3d, cv::Mat::ones(1, inl_size, points3d.type()), points3d);
} else if (test_case == TestSolver::Fundam || test_case == TestSolver::Essen) {
// create 3D points which are inliers
points3d.create(3, inl_size, pts_type);
rng.fill(points3d, cv::RNG::UNIFORM, 0.0, 1.0);
} else if (test_case == TestSolver::PnP) {
//pts1 are image points, pts2 are object points
pts2.create(3, inl_size, pts_type); // 3D inliers
rng.fill(pts2, cv::RNG::UNIFORM, 0, 1);
updateTranslation(pts2, R, t);
// project 3D points (pts2) on image plane (pts1)
pts1 = K1 * (R * pts2 + t * cv::Mat::ones(1, pts2.cols, pts2.type()));
cv::divide(pts1.row(0), pts1.row(2), pts1.row(0));
cv::divide(pts1.row(1), pts1.row(2), pts1.row(1));
// make 2D points
pts1 = pts1.rowRange(0, 2);
// create random outliers
cv::Mat pts_outliers = cv::Mat(5, out_size, pts2.type());
rng.fill(pts_outliers, cv::RNG::UNIFORM, 0, 1000);
// merge inliers with random image points = outliers
cv::hconcat(pts1, pts_outliers.rowRange(0, 2), pts1);
// merge 3D inliers with 3D outliers
cv::hconcat(pts2, pts_outliers.rowRange(2, 5), pts2);
// add Gaussian noise to image points
cv::Mat noise(pts1.rows, pts1.cols, pts1.type());
rng.fill(noise, cv::RNG::NORMAL, 0, noise_std);
pts1 += noise;
return inl_size;
} else if (test_case == TestSolver::Affine) {
} else
CV_Error(cv::Error::StsBadArg, "Unknown solver!");
if (test_case != TestSolver::PnP) {
// project 3D point on image plane
// use two relative scenes. The first camera is P1 = K1 [I | 0], the second P2 = K2 [R | t]
if (test_case != TestSolver::Affine) {
updateTranslation(points3d, R, t);
pts1 = K1 * points3d;
pts2 = K2 * (R * points3d + t * cv::Mat::ones(1, points3d.cols, points3d.type()));
// normalize by 3 coordinate
cv::divide(pts1.row(0), pts1.row(2), pts1.row(0));
cv::divide(pts1.row(1), pts1.row(2), pts1.row(1));
cv::divide(pts2.row(0), pts2.row(2), pts2.row(0));
cv::divide(pts2.row(1), pts2.row(2), pts2.row(1));
} else {
pts1 = cv::Mat(2, inl_size, pts_type);
rng.fill(pts1, cv::RNG::UNIFORM, 0, 1000);
cv::Matx33d sc(rng.uniform(1., 5.),0,0,rng.uniform(1., 4.),0,0, 0, 0, 1);
cv::Matx33d tr(1,0,rng.uniform(50., 500.),0,1,rng.uniform(50., 500.), 0, 0, 1);
const double phi = rng.uniform(0., CV_PI);
cv::Matx33d rot(cos(phi), -sin(phi),0, sin(phi), cos(phi),0, 0, 0, 1);
cv::Matx33d A = sc * tr * rot;
cv::vconcat(pts1, cv::Mat::ones(1, pts1.cols, pts1.type()), points3d);
pts2 = A * points3d;
}
// get 2D points
pts1 = pts1.rowRange(0,2); pts2 = pts2.rowRange(0,2);
// generate random outliers as 2D image points
cv::Mat pts1_outliers(pts1.rows, out_size, pts1.type()),
pts2_outliers(pts2.rows, out_size, pts2.type());
rng.fill(pts1_outliers, cv::RNG::UNIFORM, 0, 1000);
rng.fill(pts2_outliers, cv::RNG::UNIFORM, 0, 1000);
// merge inliers and outliers
cv::hconcat(pts1, pts1_outliers, pts1);
cv::hconcat(pts2, pts2_outliers, pts2);
// add normal / Gaussian noise to image points
cv::Mat noise1 (pts1.rows, pts1.cols, pts1.type()), noise2 (pts2.rows, pts2.cols, pts2.type());
rng.fill(noise1, cv::RNG::NORMAL, 0, noise_std); pts1 += noise1;
rng.fill(noise2, cv::RNG::NORMAL, 0, noise_std); pts2 += noise2;
}
return inl_size;
}
/*
* for test case = 0, 1, 2 (homography and epipolar geometry): pts1 and pts2 are 3xN
* for test_case = 3 (PnP): pts1 are 3xN and pts2 are 4xN
* all points are of the same type as model
*/
static double getError (TestSolver test_case, int pt_idx, const cv::Mat &pts1, const cv::Mat &pts2, const cv::Mat &model) {
cv::Mat pt1 = pts1.col(pt_idx), pt2 = pts2.col(pt_idx);
if (test_case == TestSolver::Homogr) { // reprojection error
// compute Euclidean distance between given and reprojected points
cv::Mat est_pt2 = model * pt1; est_pt2 /= est_pt2.at<double>(2);
if (false) {
cv::Mat est_pt1 = model.inv() * pt2; est_pt1 /= est_pt1.at<double>(2);
return (cv::norm(est_pt1 - pt1) + cv::norm(est_pt2 - pt2)) / 2;
}
return cv::norm(est_pt2 - pt2);
} else
if (test_case == TestSolver::Fundam || test_case == TestSolver::Essen) {
cv::Mat l2 = model * pt1;
cv::Mat l1 = model.t() * pt2;
if (test_case == TestSolver::Fundam) // sampson error
return fabs(pt2.dot(l2)) / sqrt(pow(l1.at<double>(0), 2) + pow(l1.at<double>(1), 2) +
pow(l2.at<double>(0), 2) + pow(l2.at<double>(1), 2));
else // symmetric geometric distance
return sqrt(pow(pt1.dot(l1),2) / (pow(l1.at<double>(0),2) + pow(l1.at<double>(1),2)) +
pow(pt2.dot(l2),2) / (pow(l2.at<double>(0),2) + pow(l2.at<double>(1),2)));
} else
if (test_case == TestSolver::PnP) { // PnP, reprojection error
cv::Mat img_pt = model * pt2; img_pt /= img_pt.at<double>(2);
return cv::norm(pt1 - img_pt);
} else
CV_Error(cv::Error::StsBadArg, "Undefined test case!");
}
/*
* inl_size -- number of ground truth inliers
* pts1 and pts2 are of the same size as from function generatePoints(...)
*/
static void checkInliersMask (TestSolver test_case, int inl_size, double thr, const cv::Mat &pts1_,
const cv::Mat &pts2_, const cv::Mat &model, const cv::Mat &mask) {
ASSERT_TRUE(!model.empty() && !mask.empty());
cv::Mat pts1 = pts1_, pts2 = pts2_;
if (pts1.type() != model.type()) {
pts1.convertTo(pts1, model.type());
pts2.convertTo(pts2, model.type());
}
// convert to homogeneous
cv::vconcat(pts1, cv::Mat::ones(1, pts1.cols, pts1.type()), pts1);
cv::vconcat(pts2, cv::Mat::ones(1, pts2.cols, pts2.type()), pts2);
thr *= 1.001; // increase a little threshold due to numerical imprecisions
const auto * const mask_ptr = mask.ptr<uchar>();
int num_found_inliers = 0;
for (int i = 0; i < pts1.cols; i++)
if (mask_ptr[i]) {
ASSERT_LT(getError(test_case, i, pts1, pts2, model), thr);
num_found_inliers++;
}
// check if RANSAC found at least 80% of inliers
ASSERT_GT(num_found_inliers, 0.8 * inl_size);
}
TEST(usac_Homography, accuracy) {
std::vector<int> gt_inliers;
const int pts_size = 1500;
cv::RNG &rng = cv::theRNG();
// do not test USAC_PARALLEL, because it is not deterministic
const std::vector<int> flags = {USAC_DEFAULT, USAC_ACCURATE, USAC_PROSAC, USAC_FAST, USAC_MAGSAC};
for (double inl_ratio = 0.1; inl_ratio < 0.91; inl_ratio += 0.1) {
cv::Mat pts1, pts2, K1, K2;
int inl_size = generatePoints(rng, pts1, pts2, K1, K2, false /*two calib*/,
pts_size, TestSolver ::Homogr, inl_ratio/*inl ratio*/, 0.1 /*noise std*/, gt_inliers);
// compute max_iters with standard upper bound rule for RANSAC with 1.5x tolerance
const double conf = 0.99, thr = 2., max_iters = 1.3 * log(1 - conf) /
log(1 - pow(inl_ratio, 4 /* sample size */));
for (auto flag : flags) {
cv::Mat mask, H = cv::findHomography(pts1, pts2,flag, thr, mask,
int(max_iters), conf);
checkInliersMask(TestSolver::Homogr, inl_size, thr, pts1, pts2, H, mask);
}
}
}
TEST(usac_Fundamental, accuracy) {
std::vector<int> gt_inliers;
const int pts_size = 2000;
cv::RNG &rng = cv::theRNG();
// start from 25% otherwise max_iters will be too big
const std::vector<int> flags = {USAC_DEFAULT, USAC_FM_8PTS, USAC_ACCURATE, USAC_PROSAC, USAC_FAST, USAC_MAGSAC};
const double conf = 0.99, thr = 1.;
for (double inl_ratio = 0.25; inl_ratio < 0.91; inl_ratio += 0.1) {
cv::Mat pts1, pts2, K1, K2;
int inl_size = generatePoints(rng, pts1, pts2, K1, K2, false /*two calib*/,
pts_size, TestSolver ::Fundam, inl_ratio, 0.1 /*noise std*/, gt_inliers);
for (auto flag : flags) {
const int sample_size = flag == USAC_FM_8PTS ? 8 : 7;
const double max_iters = 1.25 * log(1 - conf) /
log(1 - pow(inl_ratio, sample_size));
cv::Mat mask, F = cv::findFundamentalMat(pts1, pts2,flag, thr, conf,
int(max_iters), mask);
checkInliersMask(TestSolver::Fundam, inl_size, thr, pts1, pts2, F, mask);
}
}
}
TEST(usac_Fundamental, regression_19639)
{
double x_[] = {
941, 890,
596, 940,
898, 941,
894, 933,
586, 938,
902, 933,
887, 935
};
Mat x(7, 1, CV_64FC2, x_);
double y_[] = {
1416, 806,
1157, 852,
1380, 855,
1378, 843,
1145, 849,
1378, 843,
1378, 843
};
Mat y(7, 1, CV_64FC2, y_);
//std::cout << x << std::endl;
//std::cout << y << std::endl;
Mat m = cv::findFundamentalMat(x, y, USAC_MAGSAC, 3, 0.99);
EXPECT_TRUE(m.empty());
}
TEST(usac_Essential, accuracy) {
std::vector<int> gt_inliers;
const int pts_size = 1500;
cv::RNG &rng = cv::theRNG();
// findEssentilaMat has by default number of maximum iterations equal to 1000.
// It means that with 99% confidence we assume at least 34.08% of inliers
const std::vector<int> flags = {USAC_DEFAULT, USAC_ACCURATE, USAC_PROSAC, USAC_FAST, USAC_MAGSAC};
for (double inl_ratio = 0.35; inl_ratio < 0.91; inl_ratio += 0.1) {
cv::Mat pts1, pts2, K1, K2;
int inl_size = generatePoints(rng, pts1, pts2, K1, K2, false /*two calib*/,
pts_size, TestSolver ::Fundam, inl_ratio, 0.01 /*noise std, works bad with high noise*/, gt_inliers);
const double conf = 0.99, thr = 1.;
for (auto flag : flags) {
cv::Mat mask, E;
try {
E = cv::findEssentialMat(pts1, pts2, K1, flag, conf, thr, mask);
} catch (cv::Exception &e) {
if (e.code != cv::Error::StsNotImplemented)
FAIL() << "Essential matrix estimation failed!\n";
else continue;
}
// calibrate points
cv::Mat cpts1_3d, cpts2_3d;
cv::vconcat(pts1, cv::Mat::ones(1, pts1.cols, pts1.type()), cpts1_3d);
cv::vconcat(pts2, cv::Mat::ones(1, pts2.cols, pts2.type()), cpts2_3d);
cpts1_3d = K1.inv() * cpts1_3d; cpts2_3d = K1.inv() * cpts2_3d;
checkInliersMask(TestSolver::Essen, inl_size, thr / ((K1.at<double>(0,0) + K1.at<double>(1,1)) / 2),
cpts1_3d.rowRange(0,2), cpts2_3d.rowRange(0,2), E, mask);
}
}
}
TEST(usac_P3P, accuracy) {
std::vector<int> gt_inliers;
const int pts_size = 3000;
cv::Mat img_pts, obj_pts, K1, K2;
cv::RNG &rng = cv::theRNG();
const std::vector<int> flags = {USAC_DEFAULT, USAC_ACCURATE, USAC_PROSAC, USAC_FAST, USAC_MAGSAC};
for (double inl_ratio = 0.1; inl_ratio < 0.91; inl_ratio += 0.1) {
int inl_size = generatePoints(rng, img_pts, obj_pts, K1, K2, false /*two calib*/,
pts_size, TestSolver ::PnP, inl_ratio, 0.15 /*noise std*/, gt_inliers);
const double conf = 0.99, thr = 2., max_iters = 1.3 * log(1 - conf) /
log(1 - pow(inl_ratio, 3 /* sample size */));
for (auto flag : flags) {
cv::Mat rvec, tvec, mask, R, P;
CV_Assert(cv::solvePnPRansac(obj_pts, img_pts, K1, cv::noArray(), rvec, tvec,
false, (int)max_iters, (float)thr, conf, mask, flag));
cv::Rodrigues(rvec, R);
cv::hconcat(K1 * R, K1 * tvec, P);
checkInliersMask(TestSolver ::PnP, inl_size, thr, img_pts, obj_pts, P, mask);
}
}
}
TEST (usac_Affine2D, accuracy) {
std::vector<int> gt_inliers;
const int pts_size = 2000;
cv::Mat pts1, pts2, K1, K2;
cv::RNG &rng = cv::theRNG();
const std::vector<int> flags = {USAC_DEFAULT, USAC_ACCURATE, USAC_PROSAC, USAC_FAST, USAC_MAGSAC};
for (double inl_ratio = 0.1; inl_ratio < 0.91; inl_ratio += 0.1) {
int inl_size = generatePoints(rng, pts1, pts2, K1, K2, false /*two calib*/,
pts_size, TestSolver ::Affine, inl_ratio, 0.15 /*noise std*/, gt_inliers);
const double conf = 0.99, thr = 2., max_iters = 1.3 * log(1 - conf) /
log(1 - pow(inl_ratio, 3 /* sample size */));
for (auto flag : flags) {
cv::Mat mask, A = cv::estimateAffine2D(pts1, pts2, mask, flag, thr, (size_t)max_iters, conf, 0);
cv::vconcat(A, cv::Mat(cv::Matx13d(0,0,1)), A);
checkInliersMask(TestSolver::Homogr /*use homography error*/, inl_size, thr, pts1, pts2, A, mask);
}
}
}
TEST(usac_testUsacParams, accuracy) {
std::vector<int> gt_inliers;
const int pts_size = 1500;
cv::RNG &rng = cv::theRNG();
const cv::UsacParams usac_params = cv::UsacParams();
cv::Mat pts1, pts2, K1, K2, mask, model, rvec, tvec, R;
int inl_size;
auto getInlierRatio = [] (int max_iters, int sample_size, double conf) {
return std::pow(1 - exp(log(1 - conf)/(double)max_iters), 1 / (double)sample_size);
};
cv::Vec4d dist_coeff (0, 0, 0, 0); // test with 0 distortion
// Homography matrix
inl_size = generatePoints(rng, pts1, pts2, K1, K2, false, pts_size, TestSolver::Homogr,
getInlierRatio(usac_params.maxIterations, 4, usac_params.confidence), 0.1, gt_inliers);
model = cv::findHomography(pts1, pts2, mask, usac_params);
checkInliersMask(TestSolver::Homogr, inl_size, usac_params.threshold, pts1, pts2, model, mask);
// Fundamental matrix
inl_size = generatePoints(rng, pts1, pts2, K1, K2, false, pts_size, TestSolver::Fundam,
getInlierRatio(usac_params.maxIterations, 7, usac_params.confidence), 0.1, gt_inliers);
model = cv::findFundamentalMat(pts1, pts2, mask, usac_params);
checkInliersMask(TestSolver::Fundam, inl_size, usac_params.threshold, pts1, pts2, model, mask);
// Essential matrix
inl_size = generatePoints(rng, pts1, pts2, K1, K2, true, pts_size, TestSolver::Essen,
getInlierRatio(usac_params.maxIterations, 5, usac_params.confidence), 0.01, gt_inliers);
try {
model = cv::findEssentialMat(pts1, pts2, K1, K2, dist_coeff, dist_coeff, mask, usac_params);
cv::Mat cpts1_3d, cpts2_3d;
cv::vconcat(pts1, cv::Mat::ones(1, pts1.cols, pts1.type()), cpts1_3d);
cv::vconcat(pts2, cv::Mat::ones(1, pts2.cols, pts2.type()), cpts2_3d);
cpts1_3d = K1.inv() * cpts1_3d; cpts2_3d = K2.inv() * cpts2_3d;
checkInliersMask(TestSolver::Essen, inl_size, usac_params.threshold /
((K1.at<double>(0,0) + K1.at<double>(1,1) + K2.at<double>(0,0) + K2.at<double>(1,1)) / 4),
cpts1_3d.rowRange(0,2), cpts2_3d.rowRange(0,2), model, mask);
} catch (cv::Exception &e) {
if (e.code != cv::Error::StsNotImplemented)
FAIL() << "Essential matrix estimation failed!\n";
// CV_Error(cv::Error::StsError, "Essential matrix estimation failed!");
}
// P3P
inl_size = generatePoints(rng, pts1, pts2, K1, K2, false, pts_size, TestSolver::PnP,
getInlierRatio(usac_params.maxIterations, 3, usac_params.confidence), 0.01, gt_inliers);
CV_Assert(cv::solvePnPRansac(pts2, pts1, K1, dist_coeff, rvec, tvec, mask, usac_params));
cv::Rodrigues(rvec, R); cv::hconcat(K1 * R, K1 * tvec, model);
checkInliersMask(TestSolver::PnP, inl_size, usac_params.threshold, pts1, pts2, model, mask);
// P6P
inl_size = generatePoints(rng, pts1, pts2, K1, K2, false, pts_size, TestSolver::PnP,
getInlierRatio(usac_params.maxIterations, 6, usac_params.confidence), 0.1, gt_inliers);
cv::Mat K_est;
CV_Assert(cv::solvePnPRansac(pts2, pts1, K_est, dist_coeff, rvec, tvec, mask, usac_params));
cv::Rodrigues(rvec, R); cv::hconcat(K_est * R, K_est * tvec, model);
checkInliersMask(TestSolver::PnP, inl_size, usac_params.threshold, pts1, pts2, model, mask);
// Affine2D
inl_size = generatePoints(rng, pts1, pts2, K1, K2, false, pts_size, TestSolver::Affine,
getInlierRatio(usac_params.maxIterations, 3, usac_params.confidence), 0.1, gt_inliers);
model = cv::estimateAffine2D(pts1, pts2, mask, usac_params);
cv::vconcat(model, cv::Mat(cv::Matx13d(0,0,1)), model);
checkInliersMask(TestSolver::Homogr, inl_size, usac_params.threshold, pts1, pts2, model, mask);
}
}} // namespace